Anthropic shipped a skills library for lawyers.
Read it slowly.
anthropics/claude-for-legal on GitHub.
Most published work on AI agent skills is by builders, for builders. Heavy on syntax. Light on visual mental models.
The Shape of a Skill is an attempt to lower the barrier. The reader knows the shape of the territory before any single skill enters the picture.
This is a translation experiment. Use it as a map. Cartu is the method behind it.
Because trust and control sit downstream of comprehension.
The "I don't know" reaction to a new AI artifact is the barrier underneath every other adoption barrier. Naming the shape of the skill lowers the barrier to adoption and increases comprehension.
A map of the system, before any single skill enters the picture.
38 of the 151 skills follow this foundation — three skills per practice domain, each tailored to its area but architecturally identical. Learn the foundation once and you can navigate any practice domain.
Two more layers follow. Each one narrows the focus: from what's inside a single file, to the twelve archetypes those files come in.
Every skill.md in claude-for-legal carries the same skeleton — ten parts, in this order, read top to bottom. Different skills emphasize different parts; the order and the parts stay consistent. Read this once and you can navigate any skill file.
Twelve recurring shapes account for nearly all 151 skills. Each shape lights up a different combination of skeleton rows. The skeleton view shows which rows carry the weight; the drawn view shows the architectural move that gives the shape its name.
Three questions, drawn from what the catalog shows. Useful when you read a skill, useful when you write one.
Look for the What this skill does NOT do section. It sits near the end of every skill.md, and tends to run as long as the workflow body. The refusals name where the workflow's authority ends. Reading them first orients you to the skill's scope before the procedural details enter.
Look for conditions checked before any work starts, and where the skill halts. About five skills in claude-for-legal run a hard gate before intake — worker-classification is the cleanest example. That kind of gate names the condition that would make this skill the wrong tool entirely.
Every reliable skill in claude-for-legal includes a deference move — somewhere the skill stops and hands the call back. The non-lawyer consequential-action gate is one common pattern; "drafts but does not transmit" (escalation-flagger) is another. Spotting the deference move tells you where the skill's authority ends.