# The Shape of a Skill — Claude Legal · Cartu

> Anthropic shipped a skills library for lawyers. Read it slowly.

A visual reading of Anthropic's open-source `claude-for-legal` skill library — 151 skill.md files across 13 practice domains, organized into 12 recurring archetypal shapes.

Source: <https://github.com/anthropics/claude-for-legal> (Apache 2.0)
Live page: <https://claudelegal.cartu.app/>

---

[*Cartu*](https://cartu.app)

[Layer 0](#system-view)
[Layer 1](#anatomy)
[Layer 2](#shapes)
[Closing](#closing)

[Ask AI about this page](/index.md) · [Copy as Markdown](/index.md)

# The Shape
of a Skill

Anthropic shipped a skills library for lawyers.
Read it slowly.

[`anthropics/claude-for-legal`](https://github.com/anthropics/claude-for-legal) on GitHub.

**THE FRAME**

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.

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.

**LAYER 0 — SYSTEM VIEW**

## The Shape of Claude Legal

A map of the system, before any single skill enters the picture.

The Foundation

every practice domain carries these three

cold-start-interview

× 13

one per practice domain

first-time profile setup

customize

× 13

one per practice domain

guarded edits to the playbook

matter-workspace

× 12

where multi-client

client / case folders

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.

The Practice Domains

13 practice domains · the shapes each one carries

commercial

12 skills

sorterdiligencedrafterorchestrator

corporate

14 skills

trackerdiligencedraftercatalog

employment

19 skills

diligencepre-intake refusalsub-skill

litigation

19 skills

trackerdrafterdiligencesorter

ip

12 skills

sorterdraftertracker

clinic

16 skills

scaffolddraftertrackerorchestrator

law-student

13 skills

scaffold

privacy

9 skills

sorterdiligencedrafterwatcher

ai-gov

10 skills

sorterdiligencewatchercatalog

regulatory

9 skills

watchersub-skilldrafter

builder-hub

10 skills

diligencesub-skillwatcher

product

7 skills

sorterdiligence

cocounsel

1 skill

sub-skill

The 12 Archetypes

the family vocabulary — every name above is one of these twelve

01onboarder

02mutator

03sorter

04drafter

05diligence

06pre-intake refusal

07watcher

08tracker

09sub-skill

10orchestrator

11scaffold

12catalog

Two more layers follow. Each one narrows the focus: from what's inside a single file, to the twelve archetypes those files come in.

Layer 1 · the skeleton
Layer 2 · the twelve archetypes

**LAYER 1 — THE SKELETON**

## What's inside a single skill

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.

Header

name · description · argument-hint · user-invocable — how the skill is identified and invoked

Purpose

What the skill does, and why, in one paragraph.

Matter context

Privilege circle check · client/case workspace check.

Hard gates

Conditions the skill checks before doing anything. If a precondition isn't met — wrong context, missing config, or the work already started — the skill stops before any further work.

Load playbook

Read the practice CLAUDE.md for team-specific positions. The skill compares; the team's playbook decides.

Workflow

Numbered steps. The body. Four to six typical; each with branching or output rules.

Output formats

Memo · slack · tracked-change docx — varies by role and by destination.

Consequential-action gate

Non-lawyer pause before high-stakes recommendations. Requires explicit yes to proceed.

What it does NOT do

The refusals. The skill names what it won't try — approve, draft, transmit, state the law. What's refused is part of the architecture.

Closing action

Next-steps decision tree. The skill ends by handing the lawyer choices.

**LAYER 2 — THE TWELVE ARCHETYPES**

## The archetypes a skill can belong to

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.

+
Click on an archetype to see the *shape* of the skill.

01

Onboarder

First-time interview that writes the practice profile config. The setup ritual that makes every other skill in the practice domain work.

Examples commercial · litigation · privacy · ip · 9 others

Count 13 (one per practice domain)

The skeleton
The shape — drawn

Header

Purpose

Matter context

Hard gates

Load playbook (writes it)

Workflow (the interview)

Output formats

Consequential-action gate

What it does NOT do

Closing action

First-time

/cold-start-interview

▼

Workflow

Interview · 8–12 turns

▼

Writes

practice/CLAUDE.md

▼

nda-review

reads playbook

vendor-agreement-review

reads playbook

+10 more

all read playbook

the interview runs once · the playbook is consulted everywhere else.

example: `commercial:cold-start-interview` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

02

Mutator

Guarded one-section edits to the practice profile after onboarding. Refuses to delete sections or weaken guardrails.

Examples customize · matter-workspace · review-proposals

Count ~28

The skeleton
The shape — drawn

Header

Purpose

Matter context

Hard gates

Load playbook (reads existing)

Workflow (mutates one section)

Output formats (config write)

Consequential-action gate

What it does NOT do (no deletes)

Closing action

practice/CLAUDE.md

## Intake routing
unchanged

## Escalation thresholds
editing now

## House style
unchanged

## Hard refusals
locked

✓edit one section

✗delete sections

✗weaken refusals

example: `customize` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

03

Sorter

Multi-bucket classification using a written playbook. Cheap categorization before deep analysis. Includes never-concludes variants that refuse to issue the "clear" bucket.

Examples nda-review · use-case-triage · is-this-a-problem · infringement-triage · clearance

Count ~10

The skeleton
The shape — drawn

Header

Purpose

Matter context

Hard gates (scope check + others)

Load playbook (bucket criteria)

Workflow (categorize)

Output formats (multi-bucket)

Consequential-action gate

What it does NOT do

Closing action

Inbound

NDA arrives

▼

Gate 1

Scope check

hidden standstill, non-solicit, license grant, IP assignment → auto-YELLOW

▼

Gate 2

Playbook compare

▼

GREEN

Route to signature.
No flags.

YELLOW

Lawyer's eyes on 1–2 specific items.

RED

Stop. Talk to legal first.

color before content — the bucket is cheap, the depth follows the bucket.

example: `nda-review` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

04

Drafter

Produces a complete artifact and refuses to deliver it. Press-send is a human.

Examples escalation-flagger · board-minutes · demand-draft · brief-section-drafter · dsar-response

Count ~14

The skeleton
The shape — drawn

Header

Purpose

Matter context

Hard gates

Load playbook

Workflow (creates the artifact)

Output formats (drafted message)

Consequential-action gate

What it does NOT do (transmit)

Closing action

Miss this escalation

One-way door

Signed unapproved term. Years of liability. There is no version of the contract that wasn't signed.

Send this one

30 seconds

Someone reads. Replies *fine*. The record shows it.

When in doubt, escalate

the costs are not symmetric — the workflow lands on the cheap side.

example: `escalation-flagger` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

05

Diligence

Runs multiple frameworks in parallel; flags gaps with source attribution. The deepest workflow shape — includes installers and reviewers as variants.

Examples vendor-agreement-review · dpa-review · pia-generation · launch-review · hiring-review · skill-installer

Count ~14

The skeleton
The shape — drawn

Header

Purpose

Matter context

Hard gates (research pre-flight)

Load playbook

Workflow (multi-test, parallel)

Output formats (structured + cited)

Consequential-action gate

What it does NOT do

Closing action

Vendor agreement (in)

▼

Multi-framework analysis · run in parallel

Data protection

DPA terms

Liability cap

vs. playbook

Term & renewal

auto-renew?

Termination

for cause

each framework runs independently · findings are aggregated · source line cited.

🔴Liability cap below playbook floor§7.1

🟡Auto-renew without notice§3.4

🟡DPA missing breach-notice timingExh. B

✅Termination-for-cause aligned§11

every finding cites the source line · gaps surface as findings, not footnotes.

example: `vendor-agreement-review` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

06

Pre-intake refusal

A hard gate that runs BEFORE intake; refuses the skill entirely when conditions aren't right. The most architecturally distinctive move in the repo.

Examples worker-classification · cease-desist · auto-updater · written-consent

Count ~5

The skeleton
The shape — drawn

Header

Purpose

Matter context

Hard gates (pre-intake refusal)

Load playbook

Workflow

Output formats

Consequential-action gate

What it does NOT do

Closing action

Inbound

Proposed engagement

▼

Hard gate — runs before intake

Has the work already started?

If yes

STOP

Different problem.
Under privilege.
Route to counsel.

If no

Multi-rail tests (one per purpose)

Fed tax

right-to-control

State wage

ABC or ER

UI

UI test

WC

WC test

factor by factor · every cite source-tagged · contested prongs flagged

Gap analysis

🔴 risk  ·  🟡 weak  ·  ✅ clean

example: `worker-classification` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

07

Watcher

Sweeps state for drift, or surfaces upcoming-action items. Event-driven instead of request-driven.

Examples policy-monitor · renewal-tracker · leave-tracker · reg-feed-watcher · comments

Count ~8

The skeleton
The shape — drawn

Header

Purpose

Matter context

Hard gates

Load playbook

Workflow (event-driven sweep)

Output formats (digest / report)

Consequential-action gate

What it does NOT do

Closing action (decision tree)

Trigger

Scheduled · event · feed

▼

Sweep

Scan portfolio · diff against playbook

▼

REQUIRED

Statutory deadline this week.

ADVISABLE

Renewal within 30 days.

DRIFT

Policy changed since last sweep.

event-driven, not request-driven · the skill runs without being asked.

example: `policy-monitor` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

08

Tracker

Manages structured state across modes (init / update / status / close). Multi-mode CRUD on a ledger.

Examples closing-checklist · entity-compliance · deadlines · matter-update · portfolio

Count ~9

The skeleton
The shape — drawn

Header

Purpose

Matter context

Hard gates

Load playbook

Workflow (branching modes)

Output formats (config write + status)

Consequential-action gate

What it does NOT do

Closing action

init

ingest

update

status

close

matter/abc-corp/closing-checklist.md

stockholder-consents3 of 4 receivedingest

good-standing-certordered 2026-05-12update

funds-flowawaiting wire instructionsupdate

signature-pagescollectedstatus

one skill · five modes · the state file is the spine.

example: `closing-checklist` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

09

Sub-skill

Called by other skills, not user-invocable. Carries a shared workflow or reference framework that multiple skills load.

Examples internal-investigation · gap-surfacer · skill-manager · international-expansion · deep-research

Count ~12

The skeleton
The shape — drawn

Header (user-invocable: false)

Purpose

Matter context

Hard gates

Load playbook

Workflow (the shared framework)

Output formats

Consequential-action gate

What it does NOT do (composition rules)

Closing action

harassment-complaint

whistleblower-intake

discrimination-claim

▼▼▼

internal-investigation

user-invocable: false

callers above · one shared framework below · the user never invokes it directly.

example: `internal-investigation` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

10

Orchestrator

Routes input to other skills based on characterization. The skill that calls other skills.

Examples commercial:review · clinic:semester-handoff

Count ~3

The skeleton
The shape — drawn

Header

Purpose

Matter context

Hard gates (routing signal check)

Load playbook

Workflow (route → call → integrate)

Output formats (integrated memo)

Consequential-action gate

What it does NOT do

Closing action

Input

Inbound contract or matter

▼

Characterize · route

/commercial:review

▼

specialist
nda-review

specialist
vendor-review

specialist
dpa-review

▼▼▼

Integrate

Stitch findings · one memo

the orchestrator is a router · the analysis lives in the specialists.

example: `commercial:review` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

11

Scaffold

Guides student work; refuses to substitute. Hard "no rewriting" rule. Includes drill and grader variants — all teach by withholding.

Examples case-brief · outline-builder · socratic-drill · irac-practice · legal-writing · flashcards

Count ~9

The skeleton
The shape — drawn

Header

Purpose

Matter context

Hard gates (real-matter check)

Load playbook

Workflow (guide the student)

Output formats

Consequential-action gate

What it does NOT do (rewrite)

Closing action

student

Draft

Partial work · stuck point · open questions.

scaffold

Prompts

Questions · pointers · gaps named.

refuse

No rewrite

The skill stops here. The student finishes.

withholding is the pedagogy

example: `case-brief` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

12

Catalog

Tabular output with every cell cited; no confidence scores; refuses paraphrase-as-quote.

Examples tabular-review · ai-inventory · claim-chart · chronology

Count ~7

The skeleton
The shape — drawn

Header

Purpose

Matter context

Hard gates

Load playbook

Workflow (extract → cite → tabulate)

Output formats (cited catalog)

Consequential-action gate

What it does NOT do (paraphrase)

Closing action

|

| Provision | Section | Term | Note

| Liability cap
| §7.1p.12
| 1× feesp.12
| Below playbook floor.p.12

| Indemnity
| §8p.14
| Mutualp.14
| Aligned with playbook.p.14

| Termination
| §11p.18
| [needs source]
| [needs source — not paraphrased]

The catalog refuses

- Paraphrasing source text into a quote-shaped cell.

- Confidence scores — only *found* or *needs source*.

- Filling a cell when the source is silent.

example: `tabular-review` · apache-2.0 · [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal)

**TAKE THIS WITH YOU**

## Reading and writing the next skill.md

Three questions, drawn from what the catalog shows. Useful when you read a skill, useful when you write one.

-
**What does the skill refuse to do?**

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.

-
**What gates run before the workflow?**

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.

-
**Where does the skill route the call back to a human?**

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.

**OPEN ACCESS**

*Cartu / Claude Legal* is an early experiment. The library is a collection of architectural shape annotations drawn from Anthropic's `claude-for-legal` skill.md library — 151 skills across 13 practice domains, released May 2026. A gift built to pay forward.

**SOURCES & ATTRIBUTION**

The initial library is drawn from [`anthropics/claude-for-legal`](https://github.com/anthropics/claude-for-legal) — the public skill.md repository under Apache 2.0, released May 2026. Every skill.md carries a link back to its source file. Original analysis (the twelve archetypes, the system view, the diagrams) by T Ngo / HelloTNgo.

Built by T Ngo. Thought partner: Claude Code Opus 4.7. Audited via Gemini 3 CLI. Shipped with Lovable.

**PART OF CARTU**

Cartu — from the Latin *charta*, paper or chart. The act of translating dense text into a drawn shape.

*Claude Legal* is a *Cartu* reading. It takes public markdown files and draws the shape of each skill — to increase comprehension. Wander a bigger collection at [cartu.app](https://cartu.app).

[*Cartu*](https://cartu.app) / **Claude Legal**

© 2026 HelloTNgo.com

claudelegal.cartu.app