AI Field Notes from a GC
Occasional observations, real problems, and honest updates from AI educator turned GC building an AI-powered legal function.
These are previews. The full version is in the Field Guide.
Field Notes gives you the story. The Field Guide has everything else: the prompts, templates, workflows and tool reviews I use at Worksome, plus video Q&A. DKK 300 a month, with 25% off using code EXPENSEIT for as long as you stay subscribed. Cancel anytime.
I ran six real meetings through my transcript skill. It refused all six.
I built a skill that turns meeting transcripts into decisions for our knowledge base. Then I tested it on six real transcripts. It stopped on all six, and that told me more than a success would have. Here’s what it found, and the one rule I won’t compromise on: a transcript is never an answer until a person confirms it.
NDAs Are the Best Contracting Function to Automate First
A salesperson types one line into a Slack channel, and a signed NDA shows up without me writing a single email. Most NDAs should work that way. Here is the case for starting there, and the rules that keep the risky ones away from automation.
Legal AI or General AI? Start One Layer Down.
Our engineering team connected Claude to the codebase, and now we can answer product questions without pulling an engineer in. Wordsmith holds our legal and compliance knowledge. Same Slack channel. I started writing the pros and cons of keeping both and stopped halfway, because I was comparing the wrong thing. The tool is the part you can swap. Here’s what I’m building underneath it, and the question about legal I still can’t answer.
Should Your Next Legal Hire be a Legal Engineer?
I was advising an in-house legal team on their next hire. Their question: should it be a legal engineer? After our conversation, I wrote down what I actually meant by the term. It became a wheel of nine skills and ways of thinking, and a seven-question test for hiring. The catch is that one person likely won’t cover all of it.
Why I’ve been quiet
A few months ago I stopped posting as much. Field Notes went from weekly to sporadic. LinkedIn went from daily to occasional. If you noticed, you weren’t imagining it. I owe you the real reason, not the vague one.
Should you build your own legal AI tools?
Lovable lets anyone build an app without code. Here's the workflow that actually works, what I built with it, and what the security record means for you.
Before You Build in Lovable, Read This
Everyone's building in Lovable. Most are doing it wrong -- and some are exposed and don't know it yet. Three things to know before you build.
Two can't-miss Claude for Legal events next week
Next week there are two events for in-house lawyers working on AI. One is a live hands-on build session where you'll actually make something useful in under 90 minutes. One is the next in the L Suite Claude for Legal series -- with the full back catalog included when you register. Both are free. Both are designed for every skill level. Here's where to sign up.
I built a legal ops dashboard in Cowork. Here it is.
Last week I presented two live AI legal workflows to the L Suite community. A legal ops dashboard built in Cowork -- now free on GitHub. And a two-pipeline compliance system where AI builds and maintains a Notion corpus of approved product capabilities, and answers Slack questions from it automatically. No lawyer required. The recording is up. The tools are free. Here's everything from the session.
First Look: The workflow I built at a hackathon
A video walkthrough of the workflow I built at Worksome's hackathon using Claude, n8n, Notion, Slack, and Intercom.
The AI Upgrade Tax
I love technology and get jazzed when new features are released. But when frontier models ship new features, the effects don't stop at the frontier. They ripple into every legal AI tool built on top of them. Wordsmith upgraded. The workflow I had built carefully and deliberately quietly stopped working. Nothing was visibly on fire. Everything was just slightly wrong. This is the upgrade tax -- and it has taught me more about AI architecture than anything I planned to build.
The 12 Claude Legal Plugins Explained
Claude for legal plugins get to know you based on your playbook, your approvers, and your house style. Here's what each plugin is for, what the interview asks, and what changes once Claude knows you.
Installing Claude for Legal Plugins
Anthropic launched Claude for Legal — 12 practice-area plugins for in-house teams and law firms. Here's how what they are and how to install them from GitHub today.
From Chat Artifact to Command Center: What Claude Cowork Changes
I built a legal triage dashboard in Claude Chat. It worked well enough as a prototype -- until I realized I was rebuilding the context every single time I opened a new session. My friend Alex showed me what I was actually building and where it belonged. Turns out plugins, connectors, and skills are not the same thing, and knowing the difference changes how you build. This week's Field Notes is the framework. Field Guide members get the live session.
Claude for Word: The Plugin
I have strong opinions about Word plugins. I don't want to use them. So when Claude for Word launched, I assumed whatever they'd built into the plugin would eventually show up in chat too. I decided to test that assumption — no plan, just a horrible document and misplaced confidence. It failed. Badly. And I ran out of credits. The interesting part? The failure was half my fault and half Word's. I went down a rabbit hole trying to understand why, and what I found is genuinely nerdy and worth knowing if you work with documents and AI.
Two Skills + One Automation
A polite f*ck you email skill, an adversarial contract reviewer, and why I keep coming back to the same answer when people ask how to start building with AI.
Two AI skill files for in-house legal teams, a live Salesforce to Wordsmith to Slack automation in progress, and a framework for thinking about where AI actually fits in your legal function. From a GC building this in real time.
Prompts, Custom Instructions, and Skills
I asked my AI to draft a contract. It already knew my company, my format, my risk positions, and what to flag without being told. That's not a prompt. Here's what it actually is.
It's Not the Output. It's the Input.
A federal judge just ruled that privilege didn't protect a defendant's conversations with consumer Claude. Most of the coverage told you what he decided. Almost none of it told you what he left open or where the real risk actually lives. It's not the AI-assisted memo. It's what you fed in to produce it. And if you're using Claude Cowork, there's a second problem that has nothing to do with waiver and everything to do with what happens when you have to defend the privilege in court. I went to Anthropic's own support pages to understand the exposure. What I found should be in every in-house legal team's AI policy.
Claude and Wordsmith Are Making My Dashboard Dreams Come True
I've had dreams of a live, interactive legal dashboard forever. This week I started building it. I'm also covering how I restructured our Google Drive using Claude Cowork and why I never rely on just one AI tool for legal work, including the prompting technique that forces AI to argue against itself.
The File That Is a Game Changer for Claude
AI doesn't know your jurisdiction, your risk tolerance, or your edge cases. A context file fixes that. Here's what's in mine.