How a lawyer turns their own method into a Claude skill, where the free legal skill libraries stop, and why a skill you share is a skill anyone can copy.
[ KEY TAKEAWAYS ]
A Claude skill is a folder with a SKILL.md file: a name, a line on when to use it, and your instructions, with templates and checklists next to it. Claude loads it only when a task matches. The free legal skill libraries give every lawyer the same generic skills; the one worth building is your own method, written down. And a skill is plain text, so anyone you share it with can read it, copy it and pass it on.
is all a skill needs: SKILL.md, with a name, a description and instructions. Templates and checklists sit next to it.
tokens per installed skill until a task needs it. Claude reads the name and description first, the rest on demand.
tokens is Anthropic's advice for the SKILL.md body. Longer material goes into separate reference files.
of custom GPTs gave up their instructions in a 2023 Northwestern test. Text a model can read, it can be asked to repeat.
Search for Claude legal skills and you get directories: 150 here, 300 there, a Reddit thread with 160 more. They are useful the way a form book is useful. An NDA review, a clause extractor, a deposition summary, written by someone you have never met for a practice they imagined. Install a dozen and your work looks like everyone else's.
This post is about the other kind: the skill you write from the way you already work. How to build one, how it differs from a Project, and the question nobody in those directories asks, which is what happens to your method once you share it.
Four words that get used interchangeably and mean different things. Picking the right one decides who can see your work.
A procedure
Skill
How you handle a kind of task. Loads only when a task matches its description, works across conversations, and in Claude Code and the API across tools. Team and Enterprise owners can provision one for the whole organization.
loads on demand plain text
A workspace
Project
Files, instructions and chat history for one body of work. When the knowledge outgrows the context window, Claude switches to retrieval, which Anthropic says expands capacity by up to ten times.
holds documents one place
A bundle
Plugin
Skills, commands and connectors packaged together. The twelve Claude for Legal plugins are this: litigation, commercial, privacy and the rest, each a set of skills.
installs in Cowork
Someone else's server
Product
Your method running on a server, customers sending documents in and getting results back. The only shape in which you can charge for the method and keep it.
method stays private a build
The writing is the hard part, and it is the part only you can do. The format is simple.
Step 1
Write it as you would teach it
Explain the task to a new associate on paper: what you look at first, what you never skip, what makes you stop and call the client.
Step 2
Cut it to decisions
Keep the order of steps and the rules for each judgment call. Drop the background a lawyer already knows; Claude knows it too.
Step 3
Write the description last
One or two sentences on when to use it. This is what Claude reads to decide whether to load the skill, so name the documents and the moment.
Step 4
Move the long material out
Templates, clause banks and checklists go in separate files next to SKILL.md, referenced by name, so the main file stays short.
Step 5
Test on three old matters
Run it on matters you already finished and compare with what you did. Every difference is a rule you forgot to write down.
Step 6
Decide who gets it
Yourself, your firm, or the world. Each one is a different decision about who can read your method, covered below.
Here is the shape of a finished SKILL.md, using a demand letter review as the example. The header is what Claude reads first; everything under it loads when the task matches.
---
name: demand-letter-review
description: Review a personal injury demand letter before it goes out.
Use when a draft demand letter and the medical summary are attached.
---
# Demand letter review
1. Read the medical summary first. List every treatment date and provider.
2. Check that each damages figure in the letter traces to a bill in the file.
Flag any figure without a source.
3. Compare the liability paragraph against checklist.md. Never state fault
as fact if the police report is disputed.
4. Check the policy limits line against the declarations page.
5. Return: a list of fixes in order of risk, then a marked-up letter.
Reference files: checklist.md, tone-guide.mdTwo rules from Anthropic's documentation catch people out. The name is limited to 64 characters of lowercase letters, numbers and hyphens, and it cannot contain the words "claude" or "anthropic". And the description is capped at 1,024 characters, so write it like an index entry: the documents and the moment it applies. Anthropic advises keeping the body under about 5,000 tokens; a method longer than that belongs in reference files the skill points to.
The directories cover tasks every lawyer does the same way. They cannot cover the part of your work that is yours: the order you read a file in, the three things that make you turn a case down, the phrasing your best clients respond to. That is also the part that makes a skill valuable, to you and to anyone who would pay for it.
The free libraries give every lawyer the same NDA review. The skill that is worth anything is the one only you could have written.
Inside your firm, sharing is the point. On Team and Enterprise an owner can provision a skill for the whole organization so it appears for everyone, and a shared skill is view-only: colleagues can use and read it but not change it. That is exactly right for a firm's house method.
Outside the firm it is a different decision. A skill is a folder of text. Whoever receives it has your method, and can hand it to the next person. Hiding the files does not solve it either, because a model can usually be talked into repeating what it was told. When researchers at Northwestern tested custom GPTs in 2023, they extracted the system prompt from 210 of 216 (97.2%) and the uploaded files from all 24 they tried. Nobody has published the same test on Claude skills; the mechanism is the same.
Not covered by zero data retention Anthropic's documentation says Agent Skills are not covered by zero data retention arrangements. If a skill touches privileged files, put that in the firm's AI policy next to the plan's training and retention terms.
Some methods outgrow the firm. Other lawyers ask for it, a colleague at another firm wants a copy, or you realise the thing you built in evenings does a junior associate's first pass better than the junior associate. At that point the question changes from how to share it to how to charge for it without giving it away.
The answer is to keep the method on a server you control and send people the result. Customers log in, upload the documents, and get the draft back; they never see the instructions. That takes logins, billing, a server, a log of every run, and a price based on what each run actually costs in tokens. The method is the hard part, and you have done it. The rest is a build.
Other lawyers want your skill?
We turn a Claude Project or skill into a product they log into and pay for. Your method stays on the server, the accounts are yours. It starts with a free review under NDA.
A Claude skill is a folder with a SKILL.md file: a name, a short description of when to use it, and instructions, plus any templates, checklists or reference files the task needs. Claude reads only the name and description until a task matches, then loads the rest. Legal skills are skills written for legal work: reviewing an NDA against a playbook, drafting a demand letter, building a chronology. Anthropic's Claude for Legal plugins are bundles of them.
Anthropic's claude-for-legal repository on GitHub has the twelve practice-area plugins, and several community directories list between 150 and 300 free legal skills. They are a good start for common tasks. They are also everyone's: the same NDA review your opposing counsel can install. The skill that changes your work is the one written from your own method.
Write your method down as you would teach it to a new associate, cut it to the decisions and the order of steps, and put it in a file called SKILL.md under a short header with a name (lowercase letters, numbers and hyphens, up to 64 characters) and a description of when to use it (up to 1,024 characters). Put templates and checklists in separate files next to it. Upload the folder as a ZIP in the Claude app under skills, or place it in your skills folder for Claude Code. Then test it on three old matters and compare the output with what you actually did.
A Project is a workspace: files, instructions and chat history for one body of work, and it switches to retrieval when the knowledge gets bigger than the context window. A skill is a procedure: it loads only when a task needs it and works across conversations and, in Claude Code and the API, across tools. Put a matter's documents in a Project; put how you handle that kind of matter in a skill.
Yes. A skill is plain text. In the Claude app a shared skill is view-only, so colleagues can use and read it but not change it; in Claude Code it is files on disk. Even where the files are hidden, a model can usually be talked into repeating what it was told: in a 2023 Northwestern study, researchers extracted the system prompt from 210 of 216 custom GPTs and the uploaded files from all 24 they tested.
You can sell the folder, but then the buyer owns a copy of your method and can pass it on. To charge for the method without handing it over, it has to run on a server you control, with customers sending documents in and getting results back. That is a small software product: logins, billing, the server, and a log of every run.
No. Anthropic's documentation says Agent Skills are not covered by zero data retention arrangements. If a skill handles privileged material, that belongs in the firm's AI policy alongside the plan's training and retention terms.
NEWSLETTER
[ WRITTEN BY ]
Petar Jovanović
Co-Founder & Tech Lead
Co-Founder and Technical Lead at Oktopeak. Builds regulated software for legal and healthcare teams, and leads the rescues of codebases other vendors left half-finished.
[ LEGAL TECH ]
LEGAL TECH
Jun 14, 2026 · 12 min read
Claude for Legal explained for lawyers: the 12 plugins, the connectors, what it costs on each Claude plan, Clio and MyCase, and what happens to client data.
Read ArticleLEGAL TECH
Oct 9, 2026 · 8 min read
A lawyer sorted his own skills in one sentence: anything with a yes/no gate needs a person, anything that only reads can run alone. The three-question test, his skills sorted, and what the server actually changes.
Read ArticleLEGAL TECH
Mar 31, 2026 · 11 min read
You built a legal tool with Claude, Lovable or Cursor. What breaks when other firms use it, what the bar rules require, what drives the cost, and the path.
Read Article[ GET STARTED ]
30-minute call. No pitch deck. Just an honest conversation about your project.