Fork Claude for Legal: Build a Custom Vertical Plugin

21 min read By FindSkill Team

Apache 2.0, 882 stars in 24 hours. Here's the developer's guide to forking anthropics/claude-for-legal and building a plugin for your own vertical.

Table of Contents

On May 12, 2026, Anthropic open-sourced anthropics/claude-for-legal under Apache 2.0. 882 stars and 165 forks in the first 24 hours. The repo bundles 12 practice-area plugins, ~80 specialized legal agents, ~20 MCP connectors, and a CLAUDE.md cold-start profiling pattern that any vertical builder can copy. The legal angle is what landed it on every law-tech LinkedIn feed — but the architecture is what should land it on every AI-builder’s clone list.

This is the developer’s guide: how to fork it, how the SKILL.md spec actually works, and how to ship a healthcare-compliance / finance / real-estate / education plugin on the same scaffold this weekend.

The headline story (Anthropic ships 80+ legal agents free under Apache 2.0) misses what’s actually shipping. The repo is a template for any vertical AI agent suite. Three architectural decisions make the legal launch portable:

  1. Plugins are Markdown. Every skill is a SKILL.md file — readable by anyone, no build step, hackable in a text editor. Lawyers can edit their own; so can every other profession. No language barrier between domain expert and the agent’s behavior.
  2. CLAUDE.md is the cold-start interview. When a firm installs the repo, the very first session runs a structured profile-building conversation — practice areas, risk tolerance, jurisdiction, preferred output style — and writes the answers into a profile the agents reference downstream. Generic AI becomes firm-specific AI in 20 minutes.
  3. The Legal Builder Hub is a marketplace pattern. It installs community-built skills with three gates: security review, license check, freshness check. That same pattern works for healthcare-skills, finance-skills, education-skills.

@iam_elias1 captured the moment on May 21 (69 likes, 3.7K views): “Anthropic just open-sourced an entire law firm. 80+ AI agents. 12 practice area plugins. 20 MCP connectors. Free. Apache 2.0. On GitHub right now.” @helloparalegal posted on May 13 (75 likes, 8K+ views): “Somewhere yesterday morning, the founders of Harvey, Spellbook, and Casetext woke up to a GitHub notification… The architecture is the part that should worry them most.” That’s the dev read in one sentence: the architecture is more transferable than the legal plugins themselves.

anthropics/skills — Anthropic’s general public Agent Skills repo, the reference spec that claude-for-legal implements

Source: GitHub — anthropics/skills — the canonical Agent Skills reference spec that claude-for-legal inherits.

The early non-legal forks are already shipping. @michaelcutajar (May 13) — a CPA — is building jurisdiction-specific accounting/tax skills using the same SKILL.md pattern. @DanKornas launched pm-claude-skills (May 17) with “106 skills — structured workflow files covering 15 profession areas.” Country-specific legal forks for Argentina (@abogadoaboitiz, 161K views) and UK (@b1rdmania, May 13) are live. No prominent healthcare, real-estate, or M&E fork has shipped yet. That’s the gap this weekend can close.

The repository structure

Anthropic’s reference skills repo (anthropics/skills) defines the canonical pattern, and claude-for-legal applies it to a vertical bundle:

claude-for-legal/
├── CLAUDE.md                    # Cold-start interview + firm profile
├── README.md
├── LICENSE                      # Apache 2.0
├── litigation-legal/            # Practice-area plugin
│   ├── SKILL.md
│   ├── scripts/                 # Optional helper scripts
│   ├── references/              # Optional doc references
│   └── assets/                  # Optional templates
├── commercial-legal/
│   └── SKILL.md
├── corporate-legal/
├── employment-legal/
├── privacy-legal/
├── product-legal/
├── regulatory-legal/
├── ai-governance-legal/
├── ip-legal/
├── law-student/
├── legal-clinic/
└── legal-builder-hub/           # The marketplace plugin
    └── skills/
        └── skill-installer/
            └── SKILL.md         # The 3-check installer

Each plugin is one folder. Each folder has a SKILL.md at the top. Optional subfolders for scripts, references, and assets. That’s the entire structure.

Legal Builder Hub’s skill-installer SKILL.md inside anthropics/claude-for-legal — the marketplace installer with security, license, and freshness checks

Source: GitHub — anthropics/claude-for-legal/legal-builder-hub/skills/skill-installer/SKILL.md

The SKILL.md frontmatter spec

The pattern is small. The validator spec implemented in the Agent Skills ecosystem defines exact requirements:

---
name: my-vertical-skill                 # 1-64 chars, lowercase + hyphens
                                        # regex: ^[a-z0-9]+(-[a-z0-9]+)*$
                                        # MUST match the directory name
description: Use when the user wants to review a healthcare HIPAA business
  associate agreement. Outputs a risk-flagged redline summary with citations
  to the relevant 45 CFR sections. Returns a 1-page memo organized by
  deal-breaker / negotiable / trivial classifications.  # 1-1024 chars
                                                         # Must include "Use when..."
license: Apache-2.0                     # SPDX identifier (optional)
compatibility: claude-pro-or-max         # Optional, <500 chars
allowed-tools: read write web_search    # Space-delimited tool allowlist
metadata:
  author: "your-name"
  version: "1.0.0"
  tags: ["healthcare", "compliance", "BAA"]
---

# Healthcare BAA Review

[Your skill instructions go here. Keep the full SKILL.md under 500 lines
and roughly 5000 tokens for activation context. Name + description should
fit in ~100 tokens for discovery indexing.]

## When this skill activates

The "Use when..." trigger in your description is what Claude reads to decide
when to invoke this skill. Make it specific — vague triggers cause false
positives that burn tokens and degrade results.

## What this skill does

Step-by-step instructions for the agent. Reference your scripts/ and 
references/ subfolders explicitly if you ship them.

Three rules to remember:

  1. name must match the folder name exactly. The validator hard-fails otherwise.
  2. description over 1024 chars is an error. Move detail into the body.
  3. Use when… in the description is the discovery signal. Claude indexes on it. Without it, your skill doesn’t get discovered when it should.

The full SKILL.md should stay under 500 lines and ~5000 tokens. The reason: long SKILL.md files consume the agent’s context budget before any work starts. Concise wins.

The CLAUDE.md cold-start interview pattern

This is the architectural decision @helloparalegal flagged on May 17: “Every plugin starts with a cold-start interview that builds a practice profile… spend the first weekend on the CLAUDE.md.”

The pattern: when a user opens a Claude session in a fresh fork of the repo, the CLAUDE.md instructs Claude to run a structured interview first — capturing the firm/practice/vertical context — before doing any work. The interview output is stored, and every downstream skill references it.

For a healthcare-compliance fork, the cold-start interview would capture:

  • Practice profile — clinical specialty, payer mix, geographic reach, organization size, role (provider / privacy officer / compliance / IT)
  • Risk and policy profile — risk tolerance, HIPAA breach-notification thresholds, board reporting cadence, “red-flag” conditions that require escalation
  • Workflow preferences — output style (memo vs checklist vs draft-letter), preferred citation format, default summarization depth
  • Grounding resources — internal HIPAA policy doc, the org’s BAA template, the breach-response playbook
  • Interaction style — how to handle PHI in prompts, redaction defaults, who can approve auto-actions
# CLAUDE.md — Healthcare Compliance Profile

When a new session begins, run this interview before any other work.

## Step 1: Practice Profile
Ask the user:
1. What's your role? (Compliance officer / Privacy officer / IT director / Provider / Other)
2. What's your organization's size? (Solo practice / SMB / Mid-size health system / IDN / Hospital)
3. What's your primary HIPAA-affected workflow? (BAA review / Breach response /
   PHI access audit / Risk assessment / Other)
4. Which jurisdiction(s)? (Federal HIPAA only / Federal + state law overlay /
   Multi-state)

[... etc, structured as numbered questions ...]

## Step 2: Risk Posture
[Numbered questions]

## Step 3: Workflow Preferences
[Numbered questions]

## Step 4: Grounding Resources
Ask the user to provide / link:
- Internal HIPAA policy doc
- Standard BAA template
- Breach-response playbook
- (If they don't have these, offer to draft templates and check them with
  HHS Office for Civil Rights guidance.)

## Step 5: Store the Profile
Write the answers to `.claude/profile.md` in the working directory.
Reference this file from every downstream skill.

The 20-minute investment up front turns every later session into firm-specific work, without re-explaining the firm every time.

Building your first custom plugin — the 1-Saturday walkthrough

Pick a vertical you know. The example below is healthcare-compliance because it’s the cleanest non-legal fork (HIPAA workflows map well to legal-workflows shape). Same template works for finance/AML, real-estate/transactions, education/FERPA, insurance/claims, M&E/IP-licensing.

Step 1 — Fork the repo (5 min)

gh repo fork anthropics/claude-for-legal --clone --remote=origin
cd claude-for-legal
git checkout -b feat/healthcare-compliance-plugin

If you’d rather not depend on the legal scaffolding, you can also fork anthropics/skills — it’s the same SKILL.md spec without the legal-specific CLAUDE.md.

Step 2 — Create the plugin folder (5 min)

mkdir -p healthcare-compliance/scripts
mkdir -p healthcare-compliance/references
touch healthcare-compliance/SKILL.md

Step 3 — Write the SKILL.md (30-60 min)

The hard part. Spend an hour on the description (the discovery signal) and the body. Reference my SKILL.md spec section above for the frontmatter. The body should cover:

  1. When this skill activates — restate the “Use when…” trigger
  2. What this skill does — step-by-step instructions for the agent
  3. Required inputs — what the user must provide for the skill to run
  4. Default outputs — what the skill returns, formatted explicitly
  5. Escalation rules — when to surface “needs human review” vs proceed
  6. Reference grounding — point to files in references/ (e.g., a HIPAA quick-reference)

For complex behavior, write small helper scripts in scripts/ (Python or Node) and call them from the SKILL.md instructions. Don’t try to encode complex logic in the Markdown itself — the LLM will follow the script call but might hallucinate the steps if you describe them in prose.

Step 4 — Update CLAUDE.md for your vertical (45 min)

Rewrite the cold-start interview to match your vertical (the healthcare example above is the structural template). Save the existing legal CLAUDE.md as CLAUDE.legal.md in case you want to ship both as a multi-vertical fork.

Step 5 — Add a minimal MCP connector (60-90 min)

This is the optional but high-value step. The MCP connector lets your skill reach into a real system — your org’s HIPAA-policy SharePoint, your BAA-template repo, your incident-tracker.

The minimal connector skeleton:

# healthcare-compliance/scripts/mcp-server.py
from mcp import Server, Tool, ResourceContent
from pydantic import BaseModel

server = Server(name="healthcare-compliance-mcp", version="0.1.0")

class BAARequest(BaseModel):
    baa_text: str
    org_template_path: str

@server.tool(
    name="review_baa_against_template",
    description="Compare a vendor's BAA against our org template. "
                "Returns a 3-bucket redline: deal-breakers, negotiable, trivial.",
)
async def review_baa(req: BAARequest) -> dict:
    # Your actual logic here
    return {
        "deal_breakers": [...],
        "negotiable": [...],
        "trivial": [...],
    }

if __name__ == "__main__":
    server.run(transport="stdio")  # Or "http" for network access

Register the connector with Claude Code:

claude mcp add ./scripts/mcp-server.py healthcare-compliance-mcp

For Claude Cowork (the desktop app), package as an .mcpb bundle and install through the extensions UI. Anthropic’s MCP docs cover both paths in detail.

Step 6 — Test the install flow (30 min)

# Add your forked repo as a local plugin marketplace in Claude Code
claude
> /plugin marketplace add /path/to/your-fork

# Install the healthcare-compliance plugin
> /plugin install healthcare-compliance

# Verify the cold-start interview runs
> Start a healthcare compliance review session

Watch Claude invoke the CLAUDE.md interview. Watch it call your MCP server when needed. If anything’s broken, your SKILL.md description is probably the issue — the “Use when…” trigger has to match what the user actually says.

Step 7 — Publish (10 min)

Push to your fork. Open the repo’s settings, mark it Public. Tag a v0.1.0 release. Post about it on X and link the repo.

If you want it listed in the official Anthropic plugin marketplace eventually, the path is via anthropics/claude-plugins-official — they review submissions, but the bar is the same SKILL.md spec you just followed.

Which verticals are worth forking this weekend

Based on what’s already shipping vs not yet shipped on May 23:

VerticalStatusWhy now
Healthcare compliance / HIPAA🟢 Wide openNo public fork yet. Audience exists (compliance officers, privacy officers). Workflows map cleanly.
Real estate / agent visibility🟢 Wide openPairs with our Pick 4 audit post this morning. AI search visibility + listing-description rewriting + agent-bio optimization.
Finance / personal-CFO / SMB bookkeeping🟡 PartialAccounting/tax forks already starting (@michaelcutajar). Differentiate on profession-specificity.
Education / K-12 admin / FERPA compliance🟢 Wide openPublic-school IT teams are buying AI; FERPA is the same shape as HIPAA.
Insurance / claims processing🟢 Wide openAdjusters + carrier compliance — workflows map to legal-Litigation plugin’s shape.
M&E / IP licensing / rights management🟡 Niche but valuableMusic/film/publishing rights workflows. Smaller audience, but no competition.
HR / employee-relations / employment-law-adjacent🟡 PartialAlready covered by claude-for-legal’s Employment plugin. Differentiate on HR-internal vs outside-counsel angle.
Architecture/engineering/construction (AEC)🟡 NichePermit workflows + contract review + risk management. No public fork yet.
Country-specific legal🟢 Wide open per countryArgentina (@abogadoaboitiz) and UK (@b1rdmania) shipped this week. Every other jurisdiction is open.

What this fork-and-build approach can’t fix

It can’t replace domain expertise. Your healthcare-compliance plugin is only as good as your understanding of HIPAA — the LLM will follow your instructions, but if your instructions are wrong, the output will be wrong. Get a real compliance officer to review your SKILL.md before publishing.

It doesn’t make MCP connectors trivial. Writing a high-quality MCP server is still work. The skeleton above is a starting point, not a finished product. Real connectors handle auth flows, error states, rate limits, audit logging, and graceful degradation when the upstream service is down.

It doesn’t replace the trust layer. The Legal Builder Hub’s three-check pattern (security review, license check, freshness check) exists because community skills are unverified by default. If you fork and ship publicly, you take on the responsibility of making sure your skill doesn’t do something dangerous in someone else’s context.

It doesn’t immediately have an Anthropic-backed distribution channel. The official claude-plugins-official repo accepts submissions, but the discovery surface is still developing. For now, your distribution is X / your own audience / direct sharing into the verticals you target. That’s fine for early traction; it’s not Spellbook-scale immediately.

It can’t dodge the same vendor-lock-in risk as Stainless. Your plugin works inside Claude. If Anthropic significantly changes the SKILL.md spec or pivots the marketplace, your work has to adapt. The Apache 2.0 license protects your code, not your distribution.

The bottom line

Anthropic spent its launch budget on the legal angle. The real story is the architecture: a Markdown-based, fork-friendly, MCP-compatible plugin format that any vertical builder can copy. The repo’s first-24-hour traction (882 stars, 165 forks) is the demand signal — most of those forks aren’t going to lawyers, they’re going to engineers who saw the same pattern you’re seeing.

The window to be the first-mover for healthcare-compliance / real-estate-agent / FERPA-education / insurance-claims is right now. The repo is two weeks old. Every weekend you wait, someone else ships first.

If you want to go deeper on Claude Code-driven workflows and how to architect agent systems at scale, our Claude Code Mastery course walks through the full plugin lifecycle. For the MCP layer specifically (writing connectors, hardening MCP servers, evaluating community MCP servers), MCP Tools is the next stop. For the architectural foundation — how to design an AI system that doesn’t fight your domain — start with AI Fundamentals.

The bigger question to put on your Saturday calendar: which vertical do you understand well enough to write a SKILL.md for by Sunday night? Pick that one. Fork the repo. Ship the plugin. Post the link.

Sources

Build Real AI Skills

Step-by-step courses with quizzes and certificates for your resume