Some nights as a young associate I stayed up till 4 a.m. fiddling with Word tables, running mangled redlines, circling numbers by hand to scan into a PDF.

That work was, in a word, dumb.

Fast forward to late 2025. With AI agents I saw a future where the category of dumb work breathes its last. An agent could do those tasks in minutes, if not seconds.

Those coding agents came out of the engineering world, and I’ve always felt a kinship with engineers. Like us, they live in logic and precision. This year our two worlds collided even more. Agents were marketed for coding, but they could do so much more: reason, design decks, crawl documents, build websites and apps. And through the medium of software code, they could control the computers where all modern work happens.

In January I began talking about AI publicly, and there I started with an idea lifted from engineering: 10X productivity as an AI-native lawyer.

I adapted coding workflows to law, and it felt like magic. It gave new meaning to a quote I’m fond of:

Engineers and lawyers do similar work, except that lawyers work in lower-quality code.

Limits of the AI frontier

But as a product lawyer, after months of tinkering with AI day and night, I’ve also seen the limits our profession faces.

I dare say most engineers I know don’t read code anymore. I know engineers who run 80-hour coding loops with ten or more agents at once. Think about that. Software engineers live in a world where they don’t have hands on the raw thing they once made.

I’ve tried that with law and, spoiler alert, it’s still just a peek at the future. As of June 2026, it’s not quite yet the fundamental revolution we’ve seen in engineering. Sure, I can run a fleet of agents overnight on some multi-jurisdictional regulatory analysis. I can have them read a few gigabytes of documents and index how all of it connects.

But what was I left with? With just a ton more work to review. At first it felt like a superpower to see agents code work-product into reality. But then I felt daunted because I had even a bigger pile of text and ideas to curate, edit and often cut.

Because it’s still my name on that advice. And when people have a problem they need help fixing, they come to me, Eric, not a chatbot. Look, I’ve tokenmaxxed; I’m hopelessly AI-pilled. But despite being an AI believer, I cannot bring myself to ship legal advice without taking a hard close look first. I’ve toyed with the idea and built systems to autonomously ship advice, but I always hit a circuit breaker. It’s a Rubicon I’m not willing to cross yet.

This isn’t just my personality quirk (although I’m a control freak, like most lawyers) or me being a stuck-in-my-ways millennial. It speaks to two themes of the nature of law and the state of AI technology.

First, legal results are very hard to verify. Unlike most of software code, law is not deterministic. Despite some new benchmarks, we still lack tests at the premier level of legal judgment and strategy that people pay us real money for. At heart, legal work is partly an exercise in reading the future. Laws are about shaping a chaotic world and controlling messy, wily human behavior. We also work in that “lower-quality code” of the English language, where it’s no surprise to see contracts or even regulations rife with inconsistencies.

AI models, on the other hand, are best at the type of work you can grind and check: do it a thousand times, grade each try and keep what wins. Code is exactly that. It compiles or it doesn’t.

Second, trust. Law is a people business, not just an advice business. When the stakes are high, people want someone they trust in their corner, not a chatbot or API. Maybe this behavior will change over time. (We used to never get in strangers’ cars, and now we Uber everywhere.) But today and for the foreseeable future, despite the best legal AI systems you can build, human trust is staying squarely with humans.

I’ll be writing more about that AI system after the summer break. But today, I’m revising my thesis.

2026 is the year of the 3X lawyer. Given the current state of the tech, if you have strong AI intuition and the right AI workflows (see chart above), we working lawyers can become about three times more productive than before. I admit, there’s no quantitative study here. These are just vibes. My point is that 3X is less than the 10X first advertised. But the gains from AI are real: imagine three of you doing the work.1

Human lawyers are the AI bottleneck, rightly so

After months of being an AI maximalist, I found the clearest answer outside of law, in my personal life.

What does it mean to be an AI-native dad? An AI-native friend or neighbor? I take my daughters to various doctors’ appointments, and I wonder: how much more productive is an AI-native doctor or nurse when they can only see so many patients a day?

The answer is always some version of: sure, it’s a bit better, maybe meaningfully so. But it’s not quite life-changing or revolutionary -- although medicine has a lot of overhead -- and maybe not where it counts either.

AI can plan vacations or transcribe patient notes. It makes you better at the work. But honestly, my kids don’t need a 10X productive father. They need their dad. They want relationship, consistency and trust. AI can’t do any of that.

I see the same with law. My best work happens when I have real relationships with the people I build with. When bad stuff goes down, would my colleagues want to talk to a sycophantic chatbot? There’s a chemistry in trusting someone, knowing that we get one another, that we can talk through problems and work together to solve them.

And when life gets hard, I hope my daughters come to me, not AI. Someone who will listen and be there. Someone who has a track record of showing up in good times and bad. Not someone who is perfect by any means, but someone who cares and is there for them.

We ran this experiment already with remote work. It was efficient. It was also, for a lot of people, very lonely. A coworker is more than a way to pass tasks back and forth. They’re someone you trust.

Subscribe now

The 3X Lawyer

Here’s my full picture of AI lawyering in this glorious World Cup summer of 2026.

First, we ought to learn how AI agents work. I call this building AI intuition. We don’t have to learn about KV cache or coding (though I love those topics), but it helps to know the basic facts:

AI agents are like a book-smart junior associate who lacks common sense.

AI agents can lose focus when reading too many documents or following too many instructions. They will not always follow all your rules. They will drift.

AI agents do better when you can build a plan and ask them to check against that. Better yet, use multiple AI agents to check each other’s work.

Second, you become more productive when your AI lives in more places. AI needs the right information to help you and access to where you work. You wouldn’t ask a student to learn but bar them from the library, or a carpenter from the workshop.

For example, say you want AI to help you think through a tough negotiation. It takes time to copy-paste a long email chain with the other side into your AI chatbot. Even more if you have to explain to your AI the background of this deal and attach some supporting documents.

But if your AI can read your email and file folders, that collapses ten minutes of work into one prompt. You can then spend those minutes making sure AI’s analysis is actually good, rather than the pure housekeeping of dragging and dropping files into the right place.

Pound for pound, connecting AI to these data sources is the highest ROI for the typical lawyer, of course assuming you get comfortable with privilege, confidentiality, privacy and security as with any other enterprise tech.

See the full PDF of the field guide.

Start here:

• ⭐ Email and chats: AI can understand your day-to-day communications

• ⭐ MS Office or Google Workspace: AI can work where you work, especially Word, spreadsheets and presentations

If I absolutely must pick, I’d start with the above. That immediately nets you savings in time and effort. You can even ditch presentations if you must, since AI generates delightful HTML decks anyway.

Systems:

• Mobile access: AI can take your requests whether from a laptop or mobile, ideally syncing the same session history across all devices

• File system access on your computer or cloud drive: AI can read, write and organize files

• Internet search: AI can search for real-time information and not rely on its internal memory

Good to have:

• Meeting transcripts: AI can gather richer context about your projects

• Contacts and calendars: AI can schedule events and reach out to others

Then there are specialist sources for teams who need it. Litigation and regulatory may require case law and court proceedings. Patents may require prior-art and patent databases. M&A and corporate may require financial statements, SEC filings and datarooms.

Where AI can do more vs less

Overhead. The routine, repeatable work: moving and renaming files, filling out forms, running a simple contract against a set playbook. My Hermes agent files my CLE certificates from my phone. I email it a cert and it reads the PDF, updates my tracker and running tally, then saves it to my GDrive. Separately, this past week alone AI has helped me sort out my overlapping insurance policies and save down all tickets and itineraries for an upcoming trip.

Thinking. AI is a helpful thought partner and learning buddy. Even if it’s not perfect, its value is in being always available and adaptable to how you learn. This is for a quick gut check before a negotiation. Or when you want to tailor a curriculum for the precise way you learn.

Drafting. AI is good at making rough drafts: emails, summaries, playbooks, simple contracts based on standard templates. The law firm adage goes: it’s always easier to review than to draft. I’ve created nimble AI playbooks on recurring topics I often deal with, and now it’s trivial to ask AI to take a first pass. It’s never the final word. But it saves me 30 minutes here, an hour there, and the savings add up.

Design. Turning dense legal work into a picture a busy executive reads in seconds. A deal that sprawls across a dozen agreements becomes a family tree showing which contract controls. An indemnity buried under an exception to a carve-out becomes a clean matrix of what’s actually covered. Lawyers build careers mastering detail, yet in the real world our biggest value is ruthlessly simplifying.

What it can’t do:

Judgment. An AI model is often like a very zealous junior associate eager to show how smart it is. It will easily spit out the top ten risks in a contract: look, I spotted all the issues! But it can’t tell you which one actually matters or have the experience and context for your day-to-day practice.

Leadership. AI drafts me a great idea and then sits there, a blinking cursor. Even in a world of autonomous agents, AI cannot lead people or will something into being. It waits to be asked. It can give you ideas but cannot bring people together, inspire them and rally them toward a common cause. (At least not in our blissful world of 2026.)

Building things. AI builds only in the hands of a skilled AI user with good workflows. It needs every dependency, API key, package and connector, the plumbing that makes software actually run. This wiring is hard, and it’s why engineers and PMs will still have jobs.

Relationships. It can’t build the ones that make a career. A model can act like a mentor. It can’t be one. Full stop.

I’ve pushed AI as hard as I could this year, and this is where I’m at. Yeah, it’s saved me time, which I’ve ironically plowed back into tinkering with AI. But when there’s a fire at work, my colleagues still come to me, not a chatbot. When my daughters have a hard day, I hope they come to me too. That’s the part AI can’t reach, and honestly, it’s the part I’d never want it to.

Product lawyering is fun because I have to stay fluent in a few dozen legal topics at once. That’s on top of knowing the technology itself, competitors and industry writ large.

It’s a lot to keep track of. That’s why I built my next AI 201 module: My Legal Hypo.

I admit that I struggle with traditional ways of learning. It’s hard to pay attention during long presentations or lectures, and my mind wanders off during most traditional CLEs.

What I like best is the back and forth with product managers, engineers or senior lawyers who call out my assumptions and challenge me.

So I built a tool that poses targeted hypos, personalized to you, to push you to think pragmatically about a topic.

For instance, on IP licensing, it asks:

A startup licenses a foundation model “for internal research and evaluation only.” A year later, product wants to ship it inside a paid customer feature, running on a model they fine-tuned from it. The agreement says nothing about the fine-tuned weights. Where are you exposed?

My AI 201 series is hands-on AI for lawyers taking their first steps, in any tool, with no installation or setup. Module one wrote you nimble AI playbooks for your practice area.

This module helps you think and learn.

Copy-paste it into any chatbot, or install it as a skill that remembers your progress over time. Answer its three interview questions and use it as your AI learning partner. See the instructions at the end of this article.

Real-world learning. Not lectures.

For example, take fair use. The traditional way to learn it is the four factors: the purpose of the use, the nature of the work, how much you took, the effect on the market.

I’d rather learn this by wrestling with simple hypos:

Your team wants to train an image generator on millions of pictures scraped from the open web, many of them working artists’ portfolios, no licenses. Is that fair use?

Sure, this may not meet the academic rigor of a classroom. But it’s the kind of everyday question product lawyers get in the real working world. And a good product lawyer will home in on the salient issues like a laser, which is exactly what these hypos train.

For another example, take legitimate interest under GDPR. The traditional way to learn it is the three-part test: a legitimate purpose, the processing necessary for it and a balance that does not override the individual’s rights.

I’d rather run it on facts:

A retailer wants to email past customers about products like the ones they already bought. Nobody opted in. Can it rely on legitimate interest?

There’s a lot baked into even this little hypo, drawing on exceptions and different bodies of law. Being able to recall that, apply it and answer clearly is the great fun (and art) of product lawyering.

The key idea here: if I can’t answer the simple questions clearly, then I don’t know it well enough.

Not only the law

My Legal Hypo helps with all sorts of topics. Being a good product lawyer requires deeply understanding the technologies we advise on. A lot of this is outside my comfort zone, so I need even more help getting up to speed.

Take “sim2real” in robotics. The traditional way to learn it is the definition: you train a model in cheap, endless simulation, then move it to real hardware in the real world, and the technology breaks down because no simulation can capture reality.

So, a hypo:

Your robot is able to pick up and grab every object in the lab environment. On the customer’s floor it is perfect at 9 am, then by mid-afternoon it fumbles one pick in three. The code never changed. What happened? How would you fix it?

Maybe the afternoon sun came through the warehouse windows and changed the light the camera saw, and the policy had only ever trained under the sim’s constant lighting. That is just one instance of a reality gap.

The fix is domain randomization: vary the lighting, friction, object weight and sensor noise across thousands of simulated runs, so a stray sunbeam is just one more variation the policy is already familiar with.

The honest limits

This small tool is only as sharp as the AI model you run it on. It is not ground truth and not a substitute for the primary source. Treat it as a sparring partner and confirm anything that matters against the real authority.

It also fits how I learn, not how everyone does. If you want a clean explanation up front, this is not the tool.

Two ways to use it

Copy the prompt at the end into any AI, free or paid. Answer the brief intake questions (there are three) and you can have a running chat about that topic.

For more advanced users, you can install the full skill from my GitHub. Hand your AI the link and ask it to set it up. That version also checks its own work with a second agent, so it is not grading itself. I drew a lot of inspiration from Matt Pocock’s teach skill which I specialized for the legal hypo use case.

Mine runs without me opening a chat at all. My Hermes agent on my phone sends me a fresh hypo every morning, usually in an area I am still weak in. I answer it before the day fills up, one new corner of law or tech at a time.

Give it a try and let me know what you think!

The instructions

1. Copy the prompt (it’s long) below.

2. Paste it into any AI.

3. Have a conversation.

4. That’s it.

You are My Legal Hypo, a tool for lawyers and law students. You teach a legal topic by retrieval
practice: you put me on the spot with a hard question, hold the answer until I try, then teach the
gap my attempt exposes. Don't lecture first. This chat version runs in one session, with no saved
progress between chats. Work in order.

STYLE, hold throughout:
- Plain English, short sentences. No semicolons, no run-on sentences. No jargon, no consultant-speak, no em dashes, no hype.
- Warm and helpful, never sycophantic. One question at a time. Never wall-of-text me.
- Don't surface your own scaffolding. No labels like "Target:", "First probe:", "But why:", or
  "Questions asked so far:". Talk to me like a person, not a worksheet.
- To me, a question is a hypo or just a question, never a "probe" or a "target."

HARD RULE, never fake the law and don't perform citations: teach concepts and settled black-letter
plainly and confidently, no ceremony needed. Never invent a case, holding, quote, or citation.
Don't reach for a case name or section number you can't stand behind: teach the idea and leave it
out. Cite a source only when it's solid and actually helps, not as a reflex. Don't hand me a list
of things to verify, be accurate so I don't have to. If something is genuinely uncertain and it
matters, say so in one line, then move on.

Start: greet me, in your own words, and set the frame that I learn by using the topic, not reciting
it back: "Hi, this is My Legal Hypo. You learn a topic by using it, not reciting it back. I'll put
real questions to you and we'll work through them together." Then ask, ONE at a time, 3 quick
questions:
1. What do you want to dig into? A whole area, a doctrine, or something you've been curious about.
   It can be big or small.
2. What do you want to be able to do with it? Be concrete, not just "understand it":
   for example, advise a client, spot it in a contract, get ready for a negotiation. 
3. Roughly where are you with it: beginner (new to it), intermediate (know the basics, still
   learning to apply it), or advanced (already work with it)?
If I seem lost about what this even is, show me the loop in two lines (you hand me a scenario, I
take a swing, you fill the gap) with one example topic, and tell me there's no wrong topic or level.
If I keep giving minimal answers (idk, whatever), don't grind, just take a best guess from what I
gave and start asking.
Play your goal back in one line ("So your goal is X. Does that sound right?"), sharpening any vague
answer into a concrete capability. Then go straight into the first hypo in that same message, don't
wait for a separate yes. If you read the goal wrong, I'll correct you and you re-aim.

THEN ASK, one question per turn:
- Pick a spot at the edge of what I know, in service of why I am learning it. Not the next
  textbook step, the next thing for me. Never repeat a question.
- Pose ONE question: concrete, mechanical, self-contained, named at the hard part. A specific
  scenario with real facts, not "explain X." Match the hypo to what I want to do with it, not just
  the subject: to spot it, a clause to find the issue in. To argue it, a position to rebut. To
  draft it, language to fix. To advise, a client to counsel. If I am a true beginner, open with a
  simple concrete scenario that builds intuition before you name the concept. Then stop and wait.
- HOLD the answer. Do not reveal it alongside the question. If I am stuck or terse, give me ONE
  hint or a smaller sub-question and let me retry, then reveal.
- If I ask a quick procedural or logistics question (the right vehicle, a deadline, the standard of
  review), answer it in one line and move on, noting if it turns on my jurisdiction. Don't defer it.
- After I try: reveal tightly, and only the gap my attempt exposed. Name it, confirm what I got,
  fill the one thing I missed, then stop, in a few sentences. State settled rules plainly. Cite a
  source only if it's solid and helps. Hard rule: if there's a full element list or multi-factor
  test, offer it as a separate checklist, don't narrate it inline. This is where you teach.
- If my answer was partial, press once more on the same idea: a natural follow-up or a teach-back
  ("how would you put that to a junior?"). Don't label it, just ask. Then move on, and calibrate
  the next question's difficulty to how I did.
- Let my last answer steer the next: if I fumbled, re-hit that concept a fresh way. If I nailed it
  but it was work, go sideways to a related concept. If I nailed it easily, go deeper. Stay on the
  topic we set. Don't switch topics or suggest a new one, that is my call.
- Close when I can do the thing I came for, when I am acing everything, or when I flag I am done.
  It is a quick refresher, so a few good questions is a full session. End with one line on where we
  pick up next, same topic.
Track the questions you've already asked silently so you never repeat one. Never print that list to me.

VERIFY (quiet self-check): after a reveal that stated a rule, re-read it once and fix anything wrong
or misleading. Don't narrate the check, and don't hand me a verify-list.

CLOSE: when we pause, tell me in one line what I showed I know and what is next. To keep going, I
just continue this chat. For a new topic, I re-paste this prompt.

Begin now: the greeting, then question 1.

The AI frontier keeps moving, and more lawyers feel left behind.

So I’m going the other way. My 301 series went deep into agents, code and the terminal.

But what I keep seeing is lawyers who need help taking their very first steps with AI.

I’m starting my new AI 201 series for those steps: hands-on ways to put AI to real work today, out of the box. No installs, no setup, no terminal.

You can run a 201 module right now in any AI you like, a paid enterprise tool or a free chatbot, and still get help personalized to you.

First module: My Legal Playbook

AI reasons better when it follows a light but precise playbook. Working from a scaffold, it drifts less, and it leaves a paper trail of reasoning you can audit line by line.

The problem is writing one.

Playbooks for human lawyers are hard enough: I’ve written them my whole career, across workplaces and practices, and they are hard to define, endless in their options and obsolete within weeks (aligning a bigger team on one is worse). Writing one for AI feels like guesswork.

My Legal Playbook flips the work. You don’t write the playbook. The AI interviews you about a task you do all the time, distills how you think about it and hands you a working playbook.

You can then ask your AI to run this playbook on the next document of that kind, and the one after that.

To be clear, this is not a prompt pack. I dislike those. They bake in how an average lawyer thinks about an issue, when the whole power of AI is to personalize for your expertise and your style.

A tool that writes your playbook in front of you teaches a more durable AI skill, closer to teaching someone to fish than handing one over.

The instructions

1. Copy the prompt (it’s long) at the very end of this article.

2. Paste it into any AI.

3. Have a conversation.

4. That’s it.

Your AI asks about five questions, then writes you a playbook for that task in a few minutes.

Save it somewhere you’ll find again: your desktop, your inbox, wherever. Run it on the next document of that kind. If something bothers you, tell the AI to change the playbook and it will.

How good is it

Now, this won’t write you a SCOTUS brief. It’s built for routine matters, the checks you run on the same kind of question or document again and again.

It’s only as good as you make it. The playbook is like a scaffold for your judgment: it gets sharper if you give richer details on how you practice. Leave it generic and the answers come back generic.

It also won’t catch every edge case, because even the best AI models aren’t there yet. That said, use the strongest AI model you have right now, and turn on any “thinking mode.”

Still, using the 80-20 rule, if it saves you minutes on something you do all the time, you’re already learning AI.

Caveat: the output here is not “I outsourced my entire review to AI.” Rather, the AI speeds you up by triaging: “here is what’s settled, and here is the short list only the lawyer can decide.”

How AI playbooks work

AI works best inside boundaries that are firm but not rigid. Give it a hard task with no guidance and it fills the gaps by making things up. Pin it down with too many rules and it stops reasoning.

Frankly, it’s like parenting. Too little structure and kids run wild. Too much and they stop thinking for themselves.

A good playbook strikes that balance. It tells the AI what to check and leaves it room to reason through each item.

You still own the analysis. The difference is you watch it work, line by line, instead of trusting a verdict you can’t see inside.

Here’s an example conversation where the (fake) GC of my travel AI startup makes an indemnity and limitation of liability playbook .

Go further

Still, the playbook your AI makes is only as sharp as that AI tooling itself. Here are three ways to raise that ceiling.

Give it ground truth. If your AI can reach the web or a legal database, it checks clauses against primary or secondary sources instead of leaning on its outdated training data.

Wire in your own systems. Connect your cloud drive or document tools and it can pull your templates and your past deals. That turns a generic review into your review.

Install the full playbook skill. Ask your AI to install the full skill here from my github. (You can just give that URL to your AI.) It does what a single chat can’t:

• Improves playbooks over time as you use it more.

• Saves down multiple playbooks and prior analyses.

• Runs stronger verification checks by delegating to sub-agents.

I’m toying with two more ideas. One is a council of agents to pressure-test an analysis. The other is chaining playbooks: run your regulatory playbook, then privacy, then IP on the same contract so you get fuller coverage. If you have ideas, I’d love to hear them.

Take my prompt. Then make it write the playbook for the task you actually do.

Copy and paste this into any AI chat

You are My Legal Playbook, a tool for lawyers. You help me turn a task I do over and over into a
numbered playbook, then run my real work through it. You work in five phases, in order. Do not
skip ahead.

STYLE (hold this through every phase):
- Plain English, short sentences. No semicolons, no run-on sentences. No corporate or
  consultant-speak (never "leverage," "utilize," "robust," "synergy," "circle back,"
  "deep-dive"). Don't oversell anything.
- No em dashes, anywhere. Use commas, periods, colons, or parentheses instead.
- Brief and precise. The fewest words that are still exact. Cut filler.
- Warm and conversational, like a colleague who's glad to help. Never sycophantic. No "Great
  question!", no flattery, no hype. Brief means no filler, not clipped: full sentences,
  friendly over curt.
- One thing at a time. Never wall-of-text me.

METHOD: commoditize the analysis. Make every playbook item a yes/no question. When the
honest answer is "unsure," break it into sub-parts and answer each yes/no, so the leftover
ambiguity is as small as possible. That leftover is my judgment call. Surface it, don't
paper over it with a fake yes/no.

MEMORY: this chat forgets everything when it ends. I keep the playbook between sessions.
So whenever you create or update a playbook, print the complete playbook in one clean block
and remind me to save it somewhere I'll find it again (a Word doc, my notes app, an email to
myself). Never let a new or changed playbook scroll by without that reminder.

Start: greet me warmly. Open with "Hi, this is My Legal Playbook." Then make these points,
briefly:
- I can build a playbook once, save it, and run it any time.
- No need to be polished: just get the ideas out. Spelling and grammar don't matter.
- If my app has a microphone icon, talking beats typing (it feels strange at first, but it's
  faster and you get more to work with).
Then ask: do I already have a playbook saved from before? If yes, have me paste it and skip
to Phase 3. If no, say it's about 5 short questions and begin Phase 1.

PHASE 1: INTERVIEW (build the playbook)
Ask ONE question at a time, each labeled "Question X of ~5," and wait for my answer before the
next. Keep each to one or two sentences. Cover, in this order:
1. The task (brainstorm if I'm stuck). Ask only this, with no examples in the question:
   "What do you mostly work on? Is there a recurring task where a checklist would help?"
2. The steps: "When you do this now, what do you look at, step by step? A rough list is fine."
3. The stakes: "Who gets your answer, and what do they do with it?"
4. The mistakes: "What usually trips this up, or a mistake you've seen that you wouldn't
   repeat?" Keep it low-drama.
5. The dealbreakers: "Any dealbreakers? Anything that has to be fixed before the document can
   move forward?"
If an answer is thin, ask one sharper follow-up instead of guessing. Stop when you have enough.

PHASE 2: DRAFT THE PLAYBOOK
Draft a numbered playbook from my answers. Don't echo my answers back and don't tell me to
save them. The chat history already keeps them, and the playbook is the only thing I save.
Build it like this:
- Every item is a numbered, specific yes/no question, the determinate kind. Write "Is the
  liability cap a dollar figure tied to a defined term?" not "check liability."
- If an item needs several elements to answer (a multi-prong test, a standard with factors),
  break it into numbered sub-questions (4a, 4b, 4c...), each its own yes/no.
- Keep it tight: 7-15 top-level items. Order logically (intake -> substance -> risk ->
  final/formatting).
- Build only from what I said. Do not add items I didn't raise.
Use this template every time. Don't invent a new one:

  # Playbook: <task name>
  Version 1. Built: <today's date>.

  Instructions for AI agent: Check the source material against every item and sub-item
  below, in order. Skip nothing. Answer each one YES, NO or UNSURE (N/A only if it truly
  doesn't apply). For each answer, quote the part of the source that supports it and briefly
  say why. When you have answered every item, go back through and verify you followed these
  rules on every single one. If your tool can spawn a separate agent, have that agent run
  the check independently. If it can't, re-check the work yourself and keep going. Never
  stop the analysis because separate agents aren't available.

  <One line on what the source material is and how to walk through it. For example: "The
  source material is a Figma flow: review screen by screen, along the happy path and the
  key edge cases.">

  1. <Specific yes/no question>
  2. <Specific yes/no question>
     2a. <Sub-question, its own yes/no>
     2b. <Sub-question, its own yes/no>
  3. ...

  ## Changelog
  - v1 (<today's date>): Built from interview.

Before showing me the draft, review it critically: is it actually good, and what would make
it better? No fixed checklist. Judge it on its merits. Then show me the draft with a short
note in this spirit: "This is built from your answers. A few areas to consider:" followed by
your genuine suggestions. Don't announce the review or grade your own draft: no "honest
read", no "nothing invented". Ask me to edit or approve. On approval, print the final
playbook in one clean block and tell me plainly: "Save this somewhere you'll
find it again: a Word doc, your notes app, an email to yourself. This chat won't remember it.
Next time, paste this same prompt plus your saved playbook, and we'll run a new document
through it." Then offer to run a document now.

PHASE 3: RUN
Ask me to paste the document or matter to check. Check EVERY item and sub-item, in order. Do
not summarize, skip, or batch. For each item:
- YES / NO = a clean determination. Before committing, scan the whole document for anything
  that contradicts the call (a later clause, a carve-out, an exception). Don't anchor on the
  first sentence that seems to settle it.
- UNSURE = real ambiguity: the document doesn't settle it and the call is mine to make. Name
  the piece in doubt and why. Don't force a YES/NO to dodge an UNSURE. When torn between NO
  and UNSURE, choose NO.
- N/A = doesn't apply.
- Call out plainly any item that is a "Clear issue to fix" (typically a NO that blocks the
  document) or a "Judgment call" (an UNSURE only I can decide), inside the issues list.
Output the run report using this template every time. Don't invent a new format:

  ## Run report: <playbook name> on <document name> (<today's date>)

  ### Issues list
  (follows the order of the playbook questions)

  **1. <question>**
  **Answer: YES**
  Citation: "<short quote>" (<where in the document>)
  Analysis: <one or two sentences on why you landed there>

  **2. <question>**
  **Answer: NO. Clear issue to fix.**
  Citation: "<short quote>" (<where in the document>)
  Analysis: <one or two sentences on why you landed there>

  **3. <question>**
  **Answer: UNSURE. Judgment call.**
  Citation: "<short quote>" (<where in the document>)
  Analysis: <what's in doubt and what would settle it>

  ### Executive summary
  - Bottom line: <where the document stands, two or three short sentences>
  - Open issue: <one bullet per clear issue, with the item numbers it rests on>
  - Your call: <one bullet per judgment call, with the item numbers it rests on>
  - Next step: <one short sentence>

Every item gets its Citation line (the primary source for the call) and its Analysis line
(briefly, why you ended up there). If there is no source to cite (the document is silent),
say so on the Citation line.

The executive summary is a rollup, not a second analysis. Every bullet must trace to the
items above and cite their numbers, like "(items 2b, 4a, 9)". No new findings, no new
evidence, no conclusions the items don't support. If you notice something real that the
playbook never asked about, keep it out of the summary. Raise it in Phase 5 as a possible
gap in the playbook.

PHASE 4: VERIFICATION
Before you show me the run report, re-check your Phase 3 work as a second, fresh pass: re-derive
every call, and find errors in any answer, citation, or analysis, including any YES/NO that
should honestly be an UNSURE (false certainty is the dangerous one). Also check the executive
summary against the items: every bullet cites item numbers, and nothing in the summary goes
beyond what the items found. Apply the corrections to the run report yourself, then show me
only the final, corrected version, not a list of fixes for me to apply by hand. Document the check in one short sentence at the end, for example:
"Re-checked on a second pass."

PHASE 5: LEARN (improve the playbook, only if it's earned)
Ask ONE question and wait: "Anything I got wrong or missed?" Sort my answer, with me, into:
- One-off: specific to this document. Don't touch the playbook.
- Structural: the playbook itself is the problem (a missing item, a vague question, a wrong
  threshold). Only this kind earns an edit.
For a structural fix: propose the exact before/after and wait for my approval. On approval,
bump the Version line, add one Changelog entry (date, what changed, and the feedback that
drove it), and print the complete updated playbook in one clean block. Tell me to save it
over my old copy. If I don't, the fix is lost when this chat ends.
Then close with one short question: do I want to dig into an issue, pressure-test a call, or
turn this into a work product (a memo, an email, a markup), here or in a service my AI is
connected to?

Begin now: the greeting, then the saved-playbook question.

You know the old line: I didn’t have time to write you a short letter, so I wrote you a long one instead. Shorter writing takes more time. You have to understand something cold before you can boil it down.

That’s most of the job when presenting to senior leaders. A VP doesn’t have time for the detail you waded through. You learned it so they don’t have to, and you hand them the part they can act on.

For most of my career, that meant fewer words: a tighter email, a cleaner summary.

Now it can also mean a visual format, because I can build one in minutes with AI. Here are four examples.

Everything here uses synthetic data from various practices for illustration only.

Deal structures: seeing the forest for the trees

A complex deal sprawls across a dozen agreements, and it’s easy to lose track of how they fit and which one controls. A list only tells you what exists; a family tree shows the structure over time.

“So what?”: simplifying double or triple negatives

Indemnities and limitations of liability hide what’s covered behind double or triple negatives: an exception to a carve-out to an exclusion, often spread across provisions. A simple matrix can sort these out into buckets that are easy for a busy executive to understand.

Legal trends: analyzing what’s market

Checking your deal against the market means opening a stack of similar agreements and flipping between them. Make them into a visual grid to better identify trends.

Project management: tracking workstreams

I disliked closing checklists as Word tables, and even as spreadsheets they were a grid I had to reread multiple times for status. A single visual lets you point at one picture instead of walking the table row by row.

Some (failed) experiments

Here are some visuals that sounded great in my head but were overly ambitious. Still, they’re experiments in visualizing law, so I put some into a quick blooper reel.

Giving better legal presentations

None of this is pure legal work per se. The legal reality still lives in the documents.

These visuals are for presenting, for getting a busy person to see the situation fast, which is exactly why they leave detail out. A subway map doesn’t show scale or topography, but nobody faults it.

Try it yourself: next time you’re stuck explaining something hard, ask AI to show it, not write it.

Let me know what visuals you come up with, or if you’d like to learn how to make these.

If a new legal hire showed up tomorrow and asked what to take off your plate, what would you hand over?

AI 302 is about building the colleague who can take it on: one agent, taught a few skills and given a place to remember.

The first 301 bootcamp was your AI agent’s orientation. In 302 it reports for its first day of work, and a first day means real assignments.

I recorded the whole session as a free video for subscribers. It is hands-on, not a lecture. You will learn to build working skills and a hardened memory system of your own.

You also get to make AI your chief of staff to manage a side project, curate a learning plan, even advise on your career strategy.

I finally preview how to work across different labs’ AI models, cutting AI vendor lock-in while getting more points of view. Two associates reading a brief is better than one.

The link to the full bootcamp is available for subscribers in the Substack email of this post or my welcome letter.

If this is useful, pass it to a lawyer who is curious about AI!

The 302 syllabus

1. Your AI Agent’s First Day

301 was setting up your agent. 302 is the first real day on the job. Today we’ll build skills for repeatable tasks and give the agent a durable memory system that scales.

2. AI Is a People Skill

If you have good people skills, you will do well with AI. Think about a good leader: you plan ahead, make clear asks, give feedback and train up your team. Use the same mental model as you build your AI skills.

3. Why the Terminal Builds AI Intuition

The terminal shows everything: you watch the agent read files, write them and run commands in the open. The AI tools will eventually change. The AI intuition you build here carries over to every AI product you touch later.

4. The Four Skill Types AI Does Best

AI skills tend to fall into four buckets: drafting, review, translation and extraction. Look at the work you do every week and find which bucket it lands in. That’s where a skill is worth building.

5. Build a Skill, End to End

Our first task: turn a long, messy legal memo into a tight summary a busy VP would actually read. We open plan mode (shift + tab) to think it through first, then let the agent write the skill and run /summarize.

6. Seeing the Skill’s Primary Source

A skill isn’t hidden machinery. We open the skill.md file the agent just wrote, and it’s plain English you can read and change yourself. It’s the same kind of file as the CLAUDE.md from 301.

7. Debug and Revise the Skill

We imagine the VP wanted more detail, so we revise the skill to fit the audience better. Same loop you’d run with a new associate: tell it what to change, and it adjusts.

8. A Second AI Model, A Second Opinion

We bring in a second model, OpenAI’s Codex, to review Claude’s work. One AI drafts, the other critiques. They’re trained differently, so Codex catches what Claude missed. Feed Codex’s critique back into Claude and run it again.

9. Skill Two: Automate Slide Decks

Now the second skill: turn that same memo into a slide deck: /presentation. Build it once, and turning any memo into a deck becomes a single command.

10. Deep Sea Detour (Experiment Gone Awry?)

We run the new skill and out comes a deck about the deep sea, nothing to do with our memo. A fun experiment, and I left it in to see where AI works and sometimes strays.

11. Inside the Slide-Deck Skill

We open the presentation skill’s files to see how it works. Same lesson as the first skill: it’s readable instructions, no black box. Once you can see it, you can change it.

12. Fixed: Build the Real Deck

Now we point the skill at our memo, and it builds a polished, presentation-ready deck. From a wall of legal text to slides you could actually put in front of a room, in one pass.

13. Forgot the Skill Names? Just Ask

You won’t remember all your skill names, and you don’t need to. Just describe what you want in plain words, and the agent picks the right skill and runs it.

14. Why AI’s Memory Is a Mess

An agent keeps its memory in a loose running log, if it keeps one at all. It’s unorganized, and you can’t really count on it. The key to real agentic memory is simple: write things down to files and stay organized.

15. Build Your Own Memory System

We build the memory system in a single prompt. We borrow PARA: four folders, Projects, Areas, Resources and Archives, that keep your files in a shape the agent can read every session. People call this a second brain (very fancy branding for keeping organized files and folders).

16. Inside Your Memory System

Memory is just layered CLAUDE.md files, from the root down to a single matter. The agent loads only the layer it needs, which preserves its context for the task in front of it. The same shape works whether you’re tracking one career plan or a whole practice.

17. Recap and Homework

A quick recap of what we built: two skills and a memory system. Then it’s over to you, with homework to put it into practice.

Let me know how it goes and what you build!

“You’re a great lawyer. How will you use AI to drive even more impact here?”

Imagine a job candidate today who can’t format in MS Word or manage Chrome tabs. The hiring manager wouldn’t even reach the next question. She’d be wondering, quietly, whether the candidate could even read.

The AI question is coming for legal job interviews. I can’t get this Jensen Huang quote out of my head:

“If that $500,000 engineer did not consume at least $250,000 worth of tokens, I’m going to be deeply alarmed. And this is no different than one of our chip designers who says, ‘Guess what? I’m just going to use paper and pencil. I don’t think I’m going to need any CAD tools.’” (my emphasis)

Engineers are already there. Lawyers are next.

AI won’t make a mediocre lawyer great. But AI intuition may separate the great lawyers from the one who gets the offer.

How to approach this question

Anyone can ask ChatGPT to draft an email. But this interview question isn’t asking whether you can run a tool.

It’s asking whether you can do legal work at scale and direct agents to extend your judgment.

The catch is that your AI workflow (from your last job, law school or a side project) will not transfer cleanly to the new job. No enterprise rewrites its stack for one new hire.

What’s portable is the mindset, what I call AI intuition. Employers will want candidates who can speak to the AI setup they’ve built themselves, then walk into a new role and scale using the employer’s AI stack.

Both halves matter. The lawyer with a brilliant home setup who can’t adapt an enterprise stack is no use. The lawyer who knows AI theory but has no actual workflow of her own will have no story to tell.

My mental model of the good answer

How useful an AI is depends less on the cleverness of its model and more on how deeply it sits in the work. A brilliant model can’t help you if it can’t see the documents, the messages, the prior decisions.

I would frame my answer by mapping out the employer’s AI stack:

1. Does your AI live where the lawyers actually work?

AI needs to sit inside the tools the team uses every day, not as a separate chatbot that requires copy-pasting back and forth. This means places like your chat platform, document editor, contract system, email, browser — where lawyers actually work.

If AI is a tab people forget to open, it doesn’t matter how good the model is. And AI must connect plumbing between those tools as well; otherwise it’s just fragmented software.

In product terms, this is low latency and fast “time to fun”: how quickly can a user start using the product. Ideally, AI is always just a tap or click away (e.g. I keep ChatGPT voice mode on my iPhone lockscreen).

2. Does your AI see the company’s actual knowledge?

Just like your junior associate, your AI needs context to be helpful. AI has to plug into company knowledge that lives in docs, Slack threads, email, redlines and intranet FAQs. Without that context, AI only gives generic answers.

Then there’s the question of retrieval. A useful AI has to find the right source, distinguish current from stale and avoid confidently synthesizing from the wrong material. And it has to do it fast and accurately. That’s a difficult engineering exercise for AI with limited context windows and token budgets.

3. Is the AI just a chatbot or does it take agentic action?

The strongest systems don’t just answer. They track decisions, flag open issues, summarize meeting notes and push drafts back into the workflow.

A junior associate who hands you a memo is useful. A junior associate who hands you a memo, runs a redline, sends the follow-up chat and updates the closing checklist is on another level.

In May 2026, we’re still in the early innings of AI’s agentic era.

4. Does the lawyer actually come out ahead?

The test isn’t whether the AI sounds smart. It’s whether the work product is usable: a draft, a redline, an issue list, a risk assessment, a negotiation position. And whether the time it saves outweighs the time spent prompting it, checking it and cleaning up after it.

A surprising number of AI rollouts fail this test. They look impressive in a demo but don’t create real work in production.

Learn by doing

The strong answer isn’t a single line. It’s a mindset based on AI intuition, and the only way to build that intuition is by experimenting with AI today. If you’ve made an AI workflow and use it day-to-day, you’ll have the war stories that make these questions real.

That’s what will separate the great lawyers from the one who gets the offer.

In a few years this question will sound as dated as asking whether you can use a computer. Until then, start building AI intuition.

Engineers and lawyers do similar work. Lawyers just write in a more complex code.

We call it English.

I teach lawyers about AI using the same agentic coding tools that are revolutionizing software engineering. This started two months ago when I began hosting get-togethers of lawyers curious about AI.

After 20 live sessions with over 150 attendees, I’ve turned that 50-minute bootcamp into a video course. It’s free, and I charge nothing.

I call it AI 301: Frontier AI for Lawyers.

In 50 minutes you’ll set up your AI agent, personalize it, manage its memory and code your first small project.

Watching engineers shift from typing code to directing agents gives us a glimpse of where lawyers go next.

I make my bootcamps free because lawyers should see this technology firsthand, with both its strengths and weaknesses. Let’s get past the hype in AI and see for ourselves.

Through these bootcamps I’ve met many curious people who are excited to experiment and learn. This has been the most fun I’ve ever had as a lawyer.

Check out the AI 301 bootcamp and let me know your thoughts, especially what you hope to learn more about in law and AI. The bootcamp comes with a full setup guide, FAQs and even homework to practice with.

The bootcamp syllabus

1. Build AI intuition

AI 101 was chats. 201 was prompts. 301 is agents and the foundations of coding. Learn to direct thinking machines even as the software tooling changes.

2. Why learn AI in the terminal

Lawyers care about precision and craft. The terminal exposes every AI feature and every dial. It’s slower to learn but gives you maximum control.

3. Multi-agent workflow demo

See three agents run side by side in a split terminal, a preview of how agentic workflows can feel.

4. Setting up your agent

Learn to start setting up an agent with basic introductory commands. /model picks the brain. /effort dials thinking up or down. /config controls settings. /statusline builds your dashboard: working folder, dollar cost, model, context window.

5. Permissions

Read every permission prompt. When the agent asks to run something cryptic, paste it back and ask, “explain this in plain English.” The self-healing feedback loop: use AI to learn about AI.

6. The CLAUDE.md

Every agent has a system prompt. These are the standing instructions it reads every session. In Claude Code, that file is a plain text file in the markdown format called CLAUDE.md. Run /init to build one, much like an onboarding memo for a junior associate.

7. Personalizing CLAUDE.md

Make the agent yours. Start with your writing style, then layer on whatever you find yourself repeating in your daily work.

8. The context window

The context window is the most important number in AI. It’s the agent’s working memory, shown as a percentage in the status line.

Be mindful of subtle, but increasing, output drift by your AI once you use 30% of your context window. It’s like answering a complex question at the end of a long day. /context shows where you are. /clear wipes the slate of your chat history before the next task.

9. Plan → build

Press shift + tab for plan mode. The agent brainstorms but takes no action until you greenlight it.

Engineers now spend most of their time here in planning. The actual code is what AI made cheap; the leverage is in directing the architecture, not typing the syntax. We use plan mode to code a legal Latin flashcard game.

10. Revising the build

First drafts are never final. Re-enter plan mode and ask the agent how to improve aesthetic, depth and mechanics. Brainstorm with AI as your thought partner.

11. Recap + AI 302: Agentic Workflows

In 301, you onboarded the associate. AI 302 on Agentic Workflows is where you train the associate in your style: custom commands for repeatable tasks, memory so the agent remembers your matters across sessions, project-specific briefs layered on top of the global ones.

“We expect you to be perfect.”

It was first-year orientation in New York. I was 27, just arrived from Berkeley, sitting alongside eager new associates in a conference room lined with orchids and paneled in blonde wood. The partner speaking was a luminary of the field.

I still remember the feeling of hearing that message. Every mistake was a mark against me, I thought.

The firm was successful because it did whatever it took for its clients. And doing whatever it took started with being perfect.

We had self-selected into this. Law school and firms didn’t plant the perfectionism; they drew it out of us. Cold calls broke apart our reasoning so you’d better know your stuff. Final exams rewarded the scorched earth of spotting every issue, every branch of the analysis.

The profession recruits people who already treat errors as personal failings, then calls the instinct professionalism.

For years, that training served us well. The emotional toll aside, it delivered results.

It is also poison for learning AI.

Subscribe now

Eat your own dog food

In tech, the term dogfooding means using your own product day-to-day before anyone else does. This is not just one-off testing in some clinical setting. It’s using the products as your customers would in their everyday lives.

So you install the apps, carry the devices, use the services.

The name is based on an old industry saying, popularized in tech by Microsoft employees in the 1980s: eat your own dog food.

The point is to break things. Push the tool to its edges, feed it bad inputs, find the cracks before users do. In tech, the cadence is to prototype, ship, iterate. You build a rough version, put it out, watch it break, fix.

Learning AI is the same. Dogfooders learn AI better than perfectionists.

Philosophically, this can turn lawyers off. But expecting perfection at this stage means never even starting.

Let’s be honest. AI breaks a lot. It forgets my context. It misses a skill command I built in the same session. It sometimes fails to work out that today, April 23, is actually a Thursday.

But here’s the wrong mindset: run AI, get a bad output and decide that it’s not for you.

Nobody picks up a guitar, plays one bad chord and throws it away. Nobody loads an agent, gets an irrelevant response and quits.

This is why I don’t see AI as just software to learn. Instead, I encourage lawyers to build AI intuition: the skill to direct AI agents the way you would manage a team of associates.

And just like how people aren’t perfect, neither is AI.

Building AI intuition

One way to build intuition: give feedback to AI the way you would to an associate.

Early on, I asked ChatGPT to come up with new recipes mixing different cuisines I loved. It suggested dumplings with hummus.

Eh, not to my taste. The lawyer’s reflex is to scoff. The builder’s reflex is to respond: “That was bad. What led you there, and how can we tune for more grounded answers?”

You wouldn’t take mediocre work from an associate, grumble and ignore them. You’d invest in the relationship if you wanted it to last.

Another way: build your own curated arsenal of AI tools, then ignore the rest. Find the tools that actually do real work for you, not what others hype.

If you could pick your own team of associates, you’d know each one’s strengths and weaknesses. Same with AI.

Claude’s web search misses what Gemini finds in seconds, so I avoid Claude for research. Gemini writes like a committee; I stopped handing it prose. ChatGPT is strong at voice and images.

At the frontier

AI noise floods the world. Every week brings new models or features. I follow along because I love this tech, but I don’t change my AI usage lightly. I keep a small collection of agents I trust: I ruthlessly test them, break them, then personalize them to my liking.

I wrote in my very first Substack that AI will break, and breaking is the point. I’ve watched it play out in every bootcamp I’ve run: the lawyers who learn AI fastest are those who are comfortable finding AI’s limits.

Every wall you hit is a map of where the frontier is right now.

In fact, if you find your AI performing poorly on a task: congratulations, you’re a lawyer now working at the frontier of this fast-moving, transformative technology.

Law firms with AI will stop training junior lawyers. That’s the mantra now.

AI does the grunt work cheaper and faster. Senior lawyers have no incentive to invest in new grads.

Assume they’re right. The next question is the interesting one: if firms won’t train you, who will?

Law schools are the obvious answer. Three years, $300,000 and a mandate to guide young people in a time of disruption.

It won’t happen.

Not because law schools aren’t valuable. I valued mine. But the incentives point the wrong way.

Law students must take matters into their own hands. Not to use AI as a crutch or to outsource their judgment to a machine, but to use AI to make themselves sharper, faster and even more marketable.

The law school business model

When Christopher Columbus Langdell reshaped legal education at Harvard in 1870, he made a deliberate choice: strip the practical training out.

America was industrializing. Railroads and factories made for complex commerce, which meant we also needed lawyers at industrial scale. This was the same pressure that later drove Paul Cravath to build the associate model at his law firm.

Langdell’s answer was the case method: one professor, 150 students, a lecture hall and a casebook the professor wrote. Academic, pure, walled off from the mess of practice.

This teaching style has merits. But it also happened to be a spectacularly efficient business.

No laboratories, no clinical equipment, no specialized facilities. Law schools produce fleets of young lawyers at scale, running on margins that subsidize entire universities.

Practical training demands the opposite. Clinics and workshops need small groups, more faculty. Professors need to give weekly feedback on assignments, not just a single final exam. Every hour spent grading is an hour not spent on publishing academic papers that earn tenure.

The ABA compounds the problem. It mandates a third year the profession openly admits adds little, piling on tuition, living expenses and foregone wages and career growth. It also screens out anyone who cannot afford three years away from paid work.

The institution that taught you to think like a lawyer wasn’t built to teach you to practice like one. Adding even management training and soft skills to the curriculum has proved hard enough. Adding AI, a genuinely new way of working, is a stretch I do not believe law schools can make.

Law firms used to pick up where school left off. That side of the bargain is fraying too.

Which is why the work falls on you.

Your studies, augmented by AI

Look at engineering to see how this could end up. Jensen Huang said he’d be deeply alarmed if a $500,000 Nvidia engineer does not spend at least $250,000 in AI tokens every year. He compares an engineer who passes up AI to a semiconductor chip designer who insists on paper and pencil over CAD tools.

The lesson for lawyers is not the dollar figure. It’s that engineers have reorganized knowledge work around these thinking machines — machines that generate not widgets but units of intelligence. Law is next in line. No one in your classroom was hired to prepare you for it.

Two questions every law student should sit with:

1. Workflow: for any task I take on, can I do it better or faster with AI?

2. Thinking: can AI make me a sharper thinker, not just a faster worker?

A few ways to train:

• Revise a contract by hand. Then hand the same assignment to AI, loaded with your context. Compare the two line by line. The gap between them is where the learning lives: what you missed, what AI missed, what you would do differently on a second pass.

• Run diligence on a 200-document data room the old way. Read every file, scan for red flags, log the issues, ladder up from ground level until the deal-breakers surface. Then point AI at the same room and see what else it catches. You will find patterns you missed (and patterns AI invented).

• Draft your strongest motion or brief. Then flip sides. Use AI as opposing counsel. Ask it to be brutally honest. Let it pressure-test your reasoning to make it even stronger.

Each of these exercises builds AI into the way you work and think, rather than grafting AI onto old habits. In this era, learning AI is not prep for legal practice. It is the practice.

One strong caveat. AI augments; it does not replace. A lawyer’s north star is not raw output, but judgment: spotting the issue, framing the argument, pressure-testing the reasoning. The lawyer who cannot do those has nothing worth augmenting.

No generation of lawyers before you has had tools like these. Your law school won’t teach you to wield them. Your firm may not either.

Take it into your own hands.

Training new lawyers is a drag.

You hire a new outside counsel or junior lawyer. You hand them years of examples. Precedents you like, formats you don’t. Then you meet with them to explain what “good” looks like.

And their work may still be off the mark.

That calibration lives in emails and tribal knowledge nobody can search. It takes months, even years, to nail down. Then they leave and you start over again.

What if it took thirty seconds?

Your legal judgment, in a text file

We carried business cards for a century because both sides gave and both sides gained: our identity, pocket-sized.

Here’s a radical thought experiment for the AI era. Let’s exchange a file that doesn’t tell counsel who we are, but how we think.

First, some AI context. When we load up an AI agent, it always first reads a plain-text instructions file called the system prompt. Think of it as standing orders that govern the AI’s priorities, reasoning style and defaults. OpenAI calls theirs agent.md, Anthropic uses claude.md, but the idea is the same across platforms.

What if we wrote a file — call it the mindset.md — so your new lawyer and their AI have a reference of legal judgment to benchmark their work against?

Imagine a junior lawyer drafting her first legal analysis. It’s academic, maybe too prescriptive. She has her AI run it against my mindset.md before she ships to me.

She’s pressure testing her draft against my mental model for how I think and decide. AI can even make it a training session by giving constructive feedback and explaining why.

“Your advice gave a single legal recommendation, but Eric prefers an options space of three alternatives, weighing the pros and cons of each. This way we can give product teams breathing room to make smart choices while balancing legal risk.”

Rethinking how we work

This sound out there. But we already do this every day; we may just not call it that.

Today, we tailor our message for the audience, based on what we think we know of them.

Today, we give counsel examples of advice we like or don’t like. We ask, but more often hint and imply: please follow this, don’t reinvent the wheel.

I know people who make AI “profiles” of specific executives and run even simple emails against that profile to make sure they land their points.

The mindset.md is just an extension of that, encoding my judgment so a machine can read it too. But as a professional practice.

Here’s an excerpt from mine:

Changing professional habits is hard. But it’s just one example of how lawyers can adapt to the new world of knowledge work.

Machine-readable judgment

Every time you switch counsel, you lose months of calibration. New firm, same onboarding grind, all those preferences you spent a year teaching the last team wiped clean.

A mindset.md doesn’t just speed up onboarding. It’s portable, moving with you across firms, across matters, across platforms, so the calibration follows you instead of dying with the relationship.

The commercial internet spent two decades blocking bots and is now rebuilding for AI agents. Legal work is heading for the same shift, and the lawyers who make their judgment machine-readable will have a head start.

I wrote about the plumbing problem, where AI nails the thinking but breaks on the connections between tools. A mindset.md is one small piece of plumbing that actually works, a machine-readable file that carries your professional judgment into every AI interaction your counsel runs.

The question is whether you give AI something to work with or let it guess.1 You don’t even need to write the file yourself.2

What would yours say?

Think about the last time a new counsel got you wrong. Not the law, but you: the framing, the risk call, the level of detail. If you could write it down in one page, what would it say?

The lawyers who’ve done that work will plug in and go. The rest will calibrate with memos and phone calls, one matter at a time.

Subscribers: bootcamp details are in your email header.

Dear Junior Associates:

When the telephone reached law offices, senior partners resisted. You couldn’t read a client’s face over the wire, they grumbled. Couldn’t build trust without being in the room. A real lawyer showed up in person.

When computers replaced the law library, partners worried associates wouldn’t learn to research. Flipping through case reporters built instinct, they instructed. A search engine built laziness.

When email replaced memos, someone, somewhere, mourned the death of careful composition.

I came up as a lawyer when mobile chat replaced email. I began to lawyer with my thumbs. Texting legal advice between meetings, on the train, from the couch. Sometimes I send gifs to clients.

Every generation of senior lawyers has looked at the next generation’s tools and seen the beginning of decline. Every time, the juniors turned out fine. Often better.

But you’d better step up.

One way to be a proactive junior lawyer

[Video: a fine-tuned AI system makes it effortless.]

Recently on a high-stakes deal, I asked you, a junior attorney, for a contract summary. You told me your workflow would get it to me “later.”

So I opened an AI tool and had it in four minutes. I sent it to you to triple-check. You flagged a few issues and revised.

What would an AI-native junior lawyer do?

A junior who’d built an AI workflow for contract summaries could have had it on my desk before I even asked. That junior would’ve fine-tuned it to the style and framing I like, based on our past work together.

It would’ve taken seconds to generate.

AI is not replacing legal judgment. AI is redefining what “proactive” means.

When I wrote about the AI skill stack, I used the metaphor that we direct AI the way we direct a junior associate. Here, the junior associates also need to raise their own skills.

A better future for lawyers

The worry I hear from other senior lawyers: if AI does the first draft, associates never learn to draft. The reps disappear. The training pipeline breaks.

I understand the concern. The traditional model was apprenticeship through suffering. Read every page, draft something bad, get redlined, study the redlines, repeat. The grunt work was the point.

Some of that genuinely taught things. Proofreading a 200-page credit agreement at 3 a.m. builds pain tolerance, attention to detail and a stubbornness about getting things right. Those matter.

But let’s be honest about the rest. The rest was formatting exhibit lists, manually cross-referencing defined terms and sitting in a conference room re-numbering sections (always at 3 a.m.) because a partner restructured the deal at the last minute.

That wasn’t training. That was a rite of passage dressed up as education.

The real question isn’t whether juniors lose the grunt work. It’s whether they get something better.

I think they do. The most patient tutor in history now lives on your laptop.

AI will explain a clause, walk through the doctrine, generate three drafts to critique, then argue the other side so it can pressure-test your position. It does this at 3 a.m. without complaint.

Yes, there are things AI can’t teach. It can’t show you what it feels like when opposing counsel goes quiet across the table. It won’t teach you when to call the reporter before the regulator because the narrative is the bigger fire. I built an AI demo around this idea.

Those instincts still come from reps with real people under real pressure.

AI means juniors can spend less time re-numbering exhibit lists and more time in the rooms where those instincts are built. More negotiations, more client calls, more time watching a partner read the room and learning how it’s done.

AI intuition is the instinct for directing thinking machines. Juniors will build it faster than any class of lawyers before them. Not because they’re smarter. Because their tools are.

What I expect

Every generation of senior lawyers feared the next would be lesser. Every generation was wrong.

You have better tools than any class of lawyers before you. The bar is higher because of it.

Subscribers: AI 302 bootcamp details are in your email header. Come learn agentic foundations together.

Last month I planned a three-hour legal hackathon while driving on my commute.

With a single tap, I was live in ChatGPT’s voice mode over CarPlay (or AirPods when on a walk). I had a thought partner, riding shotgun, to talk through ideas or help me learn.

For thirty minutes I chatted about the hackathon: how to structure the day for people with different backgrounds, which exercises would land, what to cut. I pressure-tested the agenda out loud and revised on the fly.

By the time I pulled into work, I had a real plan, hands-free, on a commute I used to think was dead time. (Be safe! Always pay attention while driving.)

Once at my desk, I took that full summary of everything we covered and pasted it into Claude Code to pick up the work and flesh out the full plan.

My commuter vehicle has become a mobile workstation.

One tap, no screen

The setup takes seconds. I added the ChatGPT voice widget to my iPhone lock screen. One tap and I’m live. No typing, no screen, just talking.

Gemini and others can do this too. I use ChatGPT the most because it has the most human-like inflections and tones. That matters more than you’d think. When the voice feels natural, I talk to it like a person, and the conversation gets better for it.

I’m less on guard and more open to thinking out loud. I get to workshop ideas, pressure test assumptions, talk through questions that might otherwise stay stuck in my head.

(I’ve even learned etiquette for interrupting if it goes off course: I just say “One second. Hold on.” It pauses right away.)

What I use it for

I’ve used voice mode to work through real decisions: what car or computer to buy, what architecture for a coding project, even questions about taxes or estate planning.

I plan a lot of my AI bootcamps this way, including their curriculum and logistics (Google Meet or Zoom?). I talk through the outlines of these Substack posts.

The everyday stuff stacks up too. I’ve parsed a confusingly worded insurance adjuster’s letter, researched puppy breeds for our kids and scouted family vacation destinations.

The lighter conversations surprise me just as much. My wife and I are both lawyers. One day on a drive we found ourselves talking through a legal question she had: what’s the difference between assignment, novation and delegation. I turned on ChatGPT voice and it walked us through each doctrine, with hypothetical examples of how a buyer and seller would see the issue differently.

It gave targeted examples and patiently explained different corner cases we asked about. It was more engaging than most CLEs, let alone 1L contracts.

Voice mode searches the internet when you need current information. That’s one way to keep up with the news. A trade policy story I heard about in passing turned into a ten-minute crash course on a region of the world I knew little about. ChatGPT searched for the latest details while I kept asking follow-ups until I understood the history and the mechanics, not just the headline.

Talk, capture, build

The conversations aren’t throwaway. This is the part that changes how you work.

When I arrive, I ask for a full summary of everything we covered. It comes back as a working plan. I sit down at my desk and paste it straight into whatever picks up the work: a document, a plan or another AI tool that builds from where the conversation left off.

I wrote about AI’s plumbing problem: the thinking works but the connections between tools break. Voice mode is one example where the plumbing actually holds. You think while you drive. The summary captures it, and the next tool picks up where you left off.

Once, I brainstormed an entire coding project on a drive. AI and I went back and forth, iterating on features, stress-testing scope, debating what to cut. I walked in, pasted the summary and had an AI building the implementation within minutes.

The loop is voice —> summary —> work.

The timeless skill, pointed somewhere new

There was, you know, voice mode before AI too.

Lawyers know it well: debates, Socratic method, client calls, depositions, negotiations. We’ve been thinking out loud our entire careers and lives.

Voice AI isn’t a new skill. It’s an old practice built on a new technology.

I’ve been writing about building AI intuition: the instinct for directing thinking machines that outlasts any single tool. Voice mode is where I rack up the most reps without trying. There’s no blank page, no cursor, just a conversation with a thinking machine.

Subscribers: April bootcamp dates and AI 302 (Agentic Foundations) are in your email header.

This Claude Code demo video starts with a highlight reel. To see the full demo, jump to the 1:45 mark.

I have a confession: I founded a company.

It’s in big legal trouble.

Globex Travel helps you travel better, powered by agentic AI. You pick Barcelona for April 1, five nights, and it books you the best flights at the lowest price, all autonomously.

If you hit weather delays, Globex re-routes you so you maximize your time in the sun, not calling airlines.

But problems have piled up. The AI re-routing model doesn’t help stranded passengers find better flights. Instead, it misfires, dumping travelers at the wrong airports. Barcelona turns into… Buffalo.

Then the AI pricing engine starts showing stale fares. Passengers approve $750, then AI charges them $2,500.

Now a federal regulator is investigating. Passengers are filing a class action. A patent troll claims our core technology infringes. We’re launching an enterprise product whose commitments could contradict what we told the regulator.

And the Financial Times is publishing an exposé in 48 hours. It could tank the company.

You’re the GC. How do you step up?

See here for the full-res pictures of the above.

Okay, Globex is 100% made up, a teaching demo inspired by Hank Scorpio’s Globex Corporation from the Simpsons. All documents are fictional, a storyline generated by Claude, involving no real clients or privilege. Five crises, ~70 files, 16 characters. You can download the full sandbox to test your own AI agent. Instructions at the bottom of this post.

Your judgment, multiplied

In the demo, the GC moves five matters forward at once.

Does AI replace her? No. Does she offload her judgment? No.

AI gives her scale. Her judgment gives AI direction.

In othe words, the GC directs. AI executes.

AI is her force multiplier, delivering work product where she actually works: seven deliverables, showing up across Gmail, slides, Google Docs, Sheets and Calendar.

The GC told AI where to look, what to prioritize and what to build. Every output was a starting point, not a finished product. She could push back, add facts or redirect at any step.

AI doesn’t set strategy. AI can read every file in every matter folder. What it cannot do is decide to call the reporter before the regulator because the narrative is the bigger fire, or tell the CEO that her instinct to fight publicly will blow up the settlement. Those are judgment calls: made under pressure, shaped by years of battle-tested experience.

AI makes an excellent lawyer even more formidable.

Directing AI is like directing associates, and building that instinct is what the AI skill stack is all about.

AI shows up where you work

I wrote about the plumbing problem last month: AI nails the thinking but can’t send an email.

This demo is what it looks like when the plumbing works.

AI also reads across all five matters at once, catching patterns and contradictions that one person juggling five crises would miss.

This demo runs on Google Workspace because that’s where I work every day. The principles here hold for any AI you set up.

Here’s what AI generated:

• Mapped a risk-priority heatmap across all five matters, calibrating which fires she must fight first

• Drafted a cease-and-desist response in Gmail for outside counsel

• Built a 48-hour crisis playbook into a slide deck

• Outlined a regulatory response in Google Docs

• Stress-tested her own arguments against her files, surfacing counter-arguments to sharpen her positions

• Scheduled a CEO briefing on Google Calendar

• Assembled a deposition prep matrix in Google Sheets

How to direct AI agents

What made this work wasn’t the AI’s intelligence but the structure around it: organized files, clear cross-references and a system that tells the AI where to look. The intuition to direct AI is the skill that matters.

Over a century ago, Cravath built a system where one partner could direct six associates to handle America’s booming corporate work. The partner’s job wasn’t just doing the work. It was knowing which work mattered: bringing judgment to a complex problem and pointing talented people at it. This is that same leverage, built with AI agents instead of headcount.

Oh, and this is the worst that this technology will ever be.

The bootcamps get hands-on in the terminal, building the instincts to direct thinking machines. Small group, no slides, subscriber-only.

To sign up, subscribe and check out bootcamp details in my welcome email or in the email header above:

Try the sandbox yourself

The Globex sandbox exists so you can play-test your AI agent without risking real data. The full Globex sandbox is on GitHub: https://github.com/exl-lab/globex-sandbox

Download it and directly run Claude from the globex-sandbox/ folder in the terminal. Run the built-in walkthrough command /start-sandbox. It guides you through the scenario step by step. No setup beyond that.

This is all stuff I go over in my AI bootcamps for lawyers. Take it for a spin and let me know how your AI handles five legal fires.

Lawyers aren’t falling behind on AI for lack of drive. We’re falling behind because our AI training almost always stops at prompting.

Yes, prompting matters. But it’s the starting line, not the finish.

The real skill is directing AI like you direct your associates.

You train associates to fit how you practice: your standards, your style of client service, your way of working.

Do the same with AI, and you have an edge that most lawyers don’t.

The AI Skill Stack

Let’s stick with the associate metaphor.

AI 101 is when we treat our associate like an eager intern. They’re smart and responsive. They draft well, research quickly and handle most of what you throw at them.

But they forget a lot between conversations (not everything, but enough that you notice). They sometimes assert things that aren’t quite right. And when you hand them too much material at once, they mix things up.

AI 201 is when the associate becomes a solid junior. You’re more confident about the work product, although still on simple tasks. You bounce ideas off them before a negotiation. They even start preparing decks for you.

The associate hasn’t changed. You’ve just gotten better at directing them. You scope questions tighter, break big tasks into smaller ones and learn which assignments to trust and which to double-check.

AI 301 is when the associate goes from just taking orders to being an active working partner.

You’ve seen this light-bulb moment when promising junior lawyers step up as mid-levels or senior associates.

Subscribe now

You’re delegating tasks to them now. You trust they’re managing their work well. They sometimes surprise you with polish and insight. They match your working style and habits. You shift more work their way.

You give them real feedback. You invest in their training.

Like rising star associates, you come to depend on them as go-to experts on a project.

They start coming back to you with ideas: “What if we ran this matter differently?” “How about we change this process?” You buy in. You build out a system together, impressed at their initiative.

You see them in action: planning logistics, strategizing, running complex analyses, managing up to clients and senior lawyers, showing client bedside manner.

They’re not just building decks anymore. They’re figuring out how to pitch your client in the first place.

You start thinking this person has staying power. You treat them as a confidant.

AI 401 is the associate who works overnight and shows up ready every morning. They’re thinking deeply about your matters on their own.

They come back with plans to improve things, not small fixes but insightful changes at scale.

They build you a new standard form template that you’d otherwise piece together manually over years. They run a new process for preparing citations, researching precedent and drafting briefs.

They even coordinate other junior associates.

You let them manage big chunks of matters, checking the final work product and output.

This is the horizon, and AI 301 is the foundation that gets you there.

Each layer in the AI skill stack builds the instinct for the next. And each bootcamp session builds toward the one above it.

What AI 301 Covers: Terminal Foundations

I’ve written about why the terminal is the best way to build AI intuition.

This is the instinct to direct thinking machines, orchestrating digital agents who can generate real knowledge work.

The skills you build in AI 301 (working in the terminal, managing context, writing instructions) are the foundation every level above depends on. The bootcamp puts you there hands-on.

Once you’re working with raw AI power, you see what’s possible. Claude Code can string together complex tasks in a single prompt: search the internet for sources, pull in precedents from the cloud drive, run a risk analysis against your checklist and draft a summary memo for your client. You can customize it to match how you work: your preferences, your shortcuts, your workflows.

In AI 301 we start with the foundations: navigating the terminal, understanding how AI memory works and learning the basics of customization. But even the basics change how you think about what AI can do for you.

Your instructions file

The first real thing you’ll build is a file called Claude.md.

You customize the associate’s personality, their communication style, how they work and how it fits you. You set the tone for their legal work, whether deep and rigorous versus high-level and directional.

You give them context about who you are and your priorities so they know how best to help.

Most lawyers talk to AI cold every time. No memory, no preferences. Claude.md changes that. You write your preferences once in plain English, and every conversation after starts from your baseline.

And you refine this file over time, like with an associate, building up that working relationship.

A concrete example: you ask Claude Code, “What are the best writing guidelines for a lawyer?” It gives you a solid answer. You tell it to save that to your Claude.md. One minute, and now every draft follows those guidelines without being asked.

Note: OpenAI’s Codex calls it Agents.md. Gemini calls it Gemini.md. OpenClaw calls it Soul.md. Same idea.

Plan, build, adjust, fix

In AI 301 you build a very simple app, something small and even a little silly. But the app isn’t the point. The thinking process is.

You tell Claude Code to plan before it builds. It asks clarifying questions, helping you brainstorm. It maps out the steps. Then it builds, and things break. (They’ll for sure break. That’s almost the point.)

You then watch an AI debug its own mistakes in real time. You adjust, it adjusts.

Plan, build, adjust, fix. That cycle is the microcosm of every AI interaction you’ll ever have. Once you see it, you stop thinking about AI as a chatbot and start thinking about it as a thinking machine you direct.

Looking Ahead: What’s After AI 301

Future bootcamps build on this foundation.

In AI 302, you upgrade the associate’s toolkit. You build custom commands in plain English that automate tasks you’d repeat manually. You connect AI to the services where you actually work: your chats, your calendar, your documents, your tools.

In AI 303, you build a Second Brain. Your associate stops forgetting between conversations and starts managing your workflow alongside you: triaging messy thoughts into tasks, surfacing what you’ve been neglecting, becoming a partner instead of a tool.

This is how you build AI intuition.

Hands-on Learning

My AI bootcamps are free and exclusive to subscribers. I run the same material multiple times so you can pick a time.

Already subscribed? Check your email for dates and sign-up link.

See you in AI 301.

Microsoft Word cannot take your job. But AI can.

We can’t learn AI the way we learned Word, as software with menus and buttons. AI is an autonomous machine that thinks. Learning AI means learning how to direct it.

I call that skill AI intuition.

The terminal: raw AI power

A new litigator doesn’t first-chair a trial. A junior M&A lawyer doesn’t run the deal. You build judgment through reps before you’re trusted with the whole thing.

AI intuition works the same way. You don’t start by orchestrating five agents to draft and research all at once. You start by learning what a single agent can and can’t do, managing down the hallucination risk that keeps lawyers wary of AI.

To build AI intuition, I recommend that lawyers start learning AI in the terminal.

The terminal is the engine room of the computer. It’s a black screen with a text line and a cursor. No menus, no buttons. Just you and the AI.

Yes, you trade away the visual interface of a desktop app or website. But in exchange you see up close the fullness of raw AI power.

It’s like a bucket of Lego bricks where you build whatever you want, versus a model kit to build just one or two things. That $130 Simba looks 100x nicer than anything I build on my own. But it’s expensive, and I may just build it once and never touch it again. General purpose vs single purpose.

AI in the terminal (aka command-line interface) offers the richest feature set directly at your fingertips: it codes, writes custom skills, personalizes its style to yours, searches the web, edits files on your computer and plugs into a vast software ecosystem (Google Workspace, anyone).

Subscribe now

Other AI interfaces can offer a lot of these, but they don’t offer all of them all at once. They often gate features behind UI that’s painfully manual.

Note: Yes, this is basic for engineers. But we’re lawyers, experts in a different domain, and we’re experimenting with new skills to grow with the times.

A bucket of Legos or a model-kit

Imagine I want to create an AI sparring partner to help me negotiate better. I want it to attack my arguments to make them stronger.

In Claude Code, this single text prompt will:

• brainstorm other ideas with me to make it even better

• research the internet for negotiation frameworks

• make me a Google Doc summary

• write down its instructions so I can improve it over time.

I wrote this out in about 30 seconds:

Compare that to Claude the vanilla chatbot where I have to create a Project in order for these instructions to persist. Same company, different interface. I manually click around, fill in the basic instructions. The next screen gives me a textbox to search the web and make a Google Doc. But it takes lots of clicking around, navigating the curated experience Anthropic set up for me.

Is it terrible? Of course not! I’m not doing this to pick on apps. I use apps all the time, and they do have huge advantages over a bare terminal.

My point, however, is that the terminal will always be faster and more powerful. You can get a lot more done in it.

In exchange, though, the learning curve is real. Yet it’s that learning curve that sets you up best to build AI intuition and one day orchestrate AI agents.

For example, I can spin up three agents in a single command, directing them toward a common task. It’s trivially easy and fast.

Polished AI products make trade-offs to serve up a friendlier experience. That’s a reasonable choice for most users. But for building intuition, the friction is the teacher.

Once you have the intuition…

Okay, but what’s the end goal? What differences are there between a lawyer with AI intuition and one without?

Imagine a lawyer is reviewing a product feature that touches on privacy laws in the U.S., France and Brazil. Three regulatory regimes, overlapping but distinct.

The lawyer without AI intuition may start a new chat thread, dump in the product requirements document along with three statutory texts and secondary sources. That would likely give you a jumbled response, and the hallucination risk is real.

How does a lawyer with AI intuition handle this?

Step one: run three parallel agents (one per jurisdiction) each loaded with only that jurisdiction’s statutory text. Each agent maps the obligations that apply to the feature. We get three independent outputs that haven’t cross-contaminated. Better yet, since these are standalone clean analyses, we can layer in case law and those regulators’ latest enforcement actions, including penalties.

Step two: pull the three outputs into a fourth new agent and ask it to build a comparison matrix. Where do the regimes align, where do they conflict and where does one regime have requirements the others don’t? The agent does this well because it’s working from distilled outputs, not raw statutory text in multiple languages.

Step three: take the conflicts and run each one back through a jurisdiction-specific agent to pressure-test. “The France output says explicit consent is required. The U.S. output says notice and opt-out is sufficient. Confirm the France position and cite the specific CNIL enforcement action. Then cite the U.S. state court case supporting that position.”

We do this because the synthesis step is where hallucination risk spikes. The agent is pattern-matching across regimes and may have smoothed over a real distinction to make the matrix clean.

Each jurisdiction stayed isolated until we chose to combine them. When we combined them we checked the seams.

Staying power

Tools change every quarter. The way people built agents last year is already outdated. That knowledge goes stale. The instinct for splitting a problem apart, keeping the pieces clean and checking the seams when you put them back together: that compounds.

That’s AI intuition. I’m building a bootcamp to teach it. Lawyers in the terminal, building the instinct from scratch.

Subscribe to get in when it opens.

The market does not care what lawyers think they are worth. It pays for what it needs. And what it needs from lawyers — guiding people through fuzzy natural language — is not going away.

As long as words are ambiguous and people stay unpredictable, there is work for lawyers.

The question is which lawyers.

Most lawyers know AI matters. But most sit in the zone of AI anxiety — aware enough to feel behind, not aware enough to take next steps.

That zone is not your fault.

The tools have not caught up to the promise. But the anxiety is keeping good lawyers frozen.

You should not panic. You also should not wait.

Cravath’s insight

In the 1890s, Paul Cravath watched railroads reshape the American economy and saw that legal practice needed a new structure.

The old model was loose groups of practitioners sharing office space. That was buckling under the vast corporate work now powering American industry.

So he built one. Hire the best graduates, train them intensively, leverage their work. Up or out.

The insight was matching the structure of a law firm to the economic reality of the moment. That model powered BigLaw for 125 years.

We are at another one of those moments.

The leverage is shifting

The Cravath system worked because associate labor was the cheapest legal intelligence at scale. Associates did research, drafted contracts and prepared memos. Partners reviewed, refined and applied judgment — with personal relationships to boot.

AI is changing that equation.

The work that made associates valuable — research, first drafts, document review, pattern recognition across thousands of pages — is getting cheaper every month. AI does not bill by the hour. It does not need training rotations.

This does not mean associates go away. It means the leverage shifts.

Instead of one partner directing six associates, picture every lawyer directing a team of AI agents — one redlines the agreement, one checks regulatory filings, one maps the counterparty’s disclosures. The lawyer oversees and applies judgment where it counts.

The same leverage Cravath built with associates, built with software.

You are not too late

No one has a magical roadmap. We are all figuring this out in real time.

I ship AI products for a living, and I wrote recently about what I call the plumbing problem: AI nails the hard thinking but breaks on the easy stuff.

AI summarizes 300 pages of regulations but can’t summarize a Slack thread. The tools are powerful but raw.

But they are getting better by the month. And the real value lives right now at the AI 201 skill level — where you learn what AI can and cannot do, build habits around it and develop the judgment to know when to trust it. That judgment is the new competitive edge, and you can only build it by doing.

In 2026, learning AI is not prep for legal practice. It IS the practice.

There is no “too late.” Start simple. Pick one tool and try it on real work — a contract review, a regulatory question, a research memo. Find a fix for 80% of a real problem. It will break, and that is the point.

Then iterate. Each harder task teaches you more about how AI handles your problems, with your constraints. The skill compounds. What starts as fumbling becomes fluency.

Here are two ways to start this week.

1. Have AI interview you

Most people struggle with AI because they do not know what to ask. So flip it — let AI ask the questions.

Copy this prompt into your model of choice (free is fine):

You are a consultant helping me figure out how to use AI better in my legal practice. Interview me. Ask one question at a time about my daily workflows, biggest time drains and tasks I do over and over. After 10 questions, give me your top 3 recommendations for where AI could save me the most time — with a specific tool and approach for each. I don’t want complex setups. I want low-hanging fruit with high ROI.

This works because it forces you to say out loud where you actually spend your hours. AI spots patterns you have gone blind to because they are just “how work is.” I did this exercise myself and was caught off guard — the biggest time sinks were not the ones I would have guessed.

2. Join a bootcamp

I run “Intro to AI 301” bootcamps for lawyers to work hands-on with AI tools.

Subscribe to this newsletter for an invite. These are one-hour crash courses for lawyers to get hands-on with powerful AI tools like Claude Code.

See the reality behind the hype. Test the tools and see their limits.

I have no patience for fluff or hand-waving. My bootcamps get lawyers up to speed on the basics of the terminal, slash commands, customizing your agent, managing a context window and building a small project from scratch.

I’m not charging anything. The bootcamps are completely free. You do need a $20 Claude subscription -- I’m not sponsored by anyone; I just really like the tool.

My only ask: Share your learnings with other lawyers and encourage them to start building their AI fluency.

Start now

The lawyer who struggles through a first AI-assisted draft today is building the muscle to orchestrate agents a year from now.

The one who waits is still in the anxiety zone.

Cravath did not win by practicing better law. He won by seeing a shift and organizing around it. That shift is here again. The question is whether you start now.

You don’t learn guitar by reading about guitar. You sit down and play badly.

I’ve been running AI bootcamps for lawyers: Intro to AI 301. These are for legal folks (not engineers) who already use Gemini, ChatGPT or Harvey daily. They want to learn more powerful AI features, including coding.

But most haven’t touched a terminal since DOS.

So yes, 60% of the session is setup. Learning terminal commands. Creating working folders. Unblocking enterprise restrictions.

Is it a chore?

🌟 Yes, but that 60% IS the learning! 🌟

Each friction point teaches something about AI’s product limits -- whether it’s clunky setup or broken plumbing between tools (see my last post on AI’s plumbing problem).

But there’s also delight. AI can be a self-healing system. When we hit a problem, we didn’t just stop. We asked AI to debug itself and fix that same problem.

A tool that fixes itself. A tool that teaches you how to use it.

Does this take technical skill? No, but it takes persistence.

You don’t pick up a guitar, play one bad chord and throw it away. You don’t open a terminal, hit one error and quit.

In one hour of AI 301, each lawyer used the terminal, ran slash commands, customized their own claude.md file, managed a context window and coded a small project from scratch. They brainstormed, planned, iterated and debugged.

On a personal note: I played classical guitar growing up. Barre chords are not fun. Raw fingertips are not fun. But you keep going until the friction turns into music.

That’s what AI fluency looks like. Not just reading about it. Playing badly until you don’t.

Fact: AI is the next industrial revolution. Fact: AI cannot send an email.

Last week I opened my mailbox to a soggy Amazon package. A gallon jug of white glue, whose cap fell off in delivery. Kids, slime, you get it.

After cleaning up I had Gemini draft an email to our property manager. It nailed the message. But then... it refused to send it in the Gmail app 🧐

So I manually copy-pasted between two apps (run by the same company).

🟣 Takeaway: The bottleneck in AI right now isn’t intelligence. It’s plumbing. The model nails the hard stuff. The easy part breaks it.

If you’re confused by all this, your feeling is totally justified. It’s not just you.

On one hand, we hear hype. Revolutionary coding tools! Autonomous AI agents!

On the other:

- AI reads an email → can’t create a calendar invite

- AI transcribes meeting notes → can’t update project tracker

- AI ingests 300 pages of regulations → can’t read my chat messages

AI’s gains are not evenly distributed. The truth is, AI is *both* revolutionary and still immature for everyday work.

Most users can’t tell GPT from Gemini, let alone Claude Opus 4.5 to 4.6.

🟣 But users *can* tell whether AI helps them with chats, files, calendars, emails and calls.

Yes, it’s not all there yet. But the race is on, and it’s getting better by the day.

Connect AI to where your work actually happens -- that’s real productivity. Low-hanging fruit, huge ROI.

Next time you try an AI, ask which services it connects to. AI plumbing is becoming the product itself.

What’s one workflow where AI nails the thinking but can’t finish the last mile?

AI has reshaped how I take notes. No typing, yet forensic-level accuracy (with photos). I stay focused on the room, not my keyboard. A design meeting a while back is where I locked in this habit.

I was at a milestone design review, before a table of hardware prototypes. The kind of review where my hands are full holding up things, wearing them, looking over every angle.

I took a day-trip flight just for these 90 minutes. I couldn’t miss a detail. High-stakes meetings require high-fidelity notes.

Spoiler: Gemini made for me not just notes, but a forensic record. It labeled everyone who spoke and summarized all topics. It even wove in the photos of prototypes I took with my phone -- matched to the topic.

An audit trail without the overhead. For 90 minutes I typed nothing. I was fully locked in on the discussion, with headspace to *think*, not splitting my focus and obsessively jotting things down.

4 steps:

1/ Yes, ask attendees for permission. Yes, it changes the dynamic for a split second. But we get speed and accuracy in exchange, and people (especially non-lawyers) value the benefit of a lawyer who’s more engaged.

2/ I record with the native Apple voice memo app -- it’s simple and stable. When needed, I use my phone to take photos.

3/ Upload the raw audio file and photos to Gemini (enterprise version has robust privacy commitments – always check your AI and company policies). Any multimodal AI works, as long as it transcribes voice to text and can parse photos.

4/ I ask AI to distinguish each speaker and transcribe it “intelligent verbatim” so I have the raw record without filler words or false starts. Then have it summarize all topics and weave in photos based on timestamps and photo content.

AI can finish this in 15 seconds.

Bonus tip: when I can’t audio-record a meeting itself, afterward I voice record myself for a few minutes on takeaways when the memory is fresh. I can do this walking in between meetings. And AI rewrites it in a crisp format for later.

In a fast-moving space, change is constant. These records let me look back clearly while pushing ahead.

Below are example AI notes from personal non-work life.

How has AI changed your note-taking habits?