Claude for Legal Claude skills

Setting up a plugin: what the cold-start interview does

Each practice-area plugin ships its own cold-start-interview skill, and the other skills in that plugin will not run from an empty profile. If the profile is missing or still contains [PLACEHOLDER] markers, a skill such as ai-inventory sends you back to the interview first. The AI governance plugin runs the interview in this order:

1. It checks whether ~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md exists. If the file is populated and you did not pass --redo, it asks before overwriting.
2. It runs the interview, starting with a Part 0 role and integration check.
3. It asks for seed documents: your AI or acceptable use policy (URL or file), a prior impact assessment, key vendor AI agreements, and a model inventory or allowlist/blocklist if you have one.
4. It extracts policy commitments and prohibitions, vendor positions with gaps noted, the structure of your impact assessments, and approved or prohibited tool lists.
5. If a populated profile exists under the plugin cache path (~/.claude/plugins/cache/claude-for-legal/ai-governance-legal/*/CLAUDE.md) but not at the config path, it copies the profile over and tells you what it migrated.
6. It writes the profile, shows a summary and offers a first task.

The skill has a rule against writing a profile with silent gaps. When you skip a question, the profile records a placeholder that you chose to leave, so a question cannot get lost in the scrollback. Collect your seed documents before you start. The cold-start skills for commercial and AI governance are the largest files in the sample, at roughly 47 to 49 KB each, so expect a long first session.

You will use two flags after setup. --redo reruns the full interview and overwrites the profile. --check-integrations rescans available MCP connectors and refreshes only the ## Available integrations table. Run it after you connect Slack, document storage or scheduled tasks. A connector gets a ✓ only after an MCP tool call actually succeeds. A connector that is configured but untested shows ⚪ with a one-line note on how to confirm it. A .mcp.json declaration by itself does not earn a ✓.

The sampled plugins also include a customize skill and a matter-workspace skill. Matter workspaces are off by default for in-house users, and in that mode skills work from practice-level context only. When you turn them on, skills ask which matter you are working on (/ai-governance-legal:matter-workspace switch <slug>), load that matter's matter.md, and write outputs to ~/.claude/plugins/config/claude-for-legal/ai-governance-legal/matters/<matter-slug>/. A skill does not read another matter's files unless Cross-matter context is on. Firms and clinics that need to keep client work separate should check this setting.

Slash commands and agents by plugin

In the README, each named agent maps to one plugin command. These are the commands the README lists:

Renewal Watcher, Deal Debrief and Playbook Monitor in commercial-legal have no slash command. They run as scheduled agents. Renewal Watcher scans the contract register for cancel-by and renewal deadlines. Deal Debrief does a weekly sweep of signed agreements that deviated from the playbook and asks the attorney to log context while it is still fresh. Playbook Monitor watches that deviation log and proposes playbook updates when a clause has drifted. The README treats these and the docket, regulatory feed, diligence grid and launch radar watchers as managed-agent cookbooks.

The README also lists plugins with no agent table entry in the excerpt. ip-legal covers trademark clearance, FTO triage, cease-and-desist drafting and triage, DMCA takedowns and counter-notices, and portfolio tracking. legal-clinic covers professor setup, a student semester ramp, supervisor guides with an assist, guide or teach posture, and structured intake with cross-area issue spotting.

To get started, pick the agent closest to your daily work, run that plugin's interview, and then adjust the underlying skill, the profile and the connectors, as the README suggests.

Example workflow: AI impact assessment and EU AI Act inventory

The AI governance skills show how a typical run goes. The impact assessment skill takes a short description of the use case:

/ai-governance-legal:aia-generation "AI résumé screening for HR"

The skill first reads the practice profile and confirms that your impact assessment house style is populated. Next it picks a fast or full risk track based on governance tier and the characteristics of the use case. Intake is a conversation, not a form. For each regime in your footprint, the skill researches the risk tier, prohibited-practice exposure and applicable obligations, citing primary sources. It warns against assuming that internal-only systems are out of scope, because most regimes treat employee data as personal data and employee monitoring as consequential. The assessment is written in your house style (from your seed document, or a default if you provided none) and then checked against the AI policy commitments in your profile. You get the assessment document, a list of conditions, and handoff flags such as a privacy PIA or a vendor review.

/ai-governance-legal:ai-inventory keeps the register that feeds this work. It takes list, add, edit <id>, classify <id> or show <id> and stores records in ~/.claude/plugins/config/claude-for-legal/ai-governance-legal/ai-systems.yaml. The skill records role (provider, deployer, importer, distributor, authorized representative, product manufacturer) and risk tier (prohibited, high-risk, limited, minimal, GPAI, GPAI with systemic risk) for each system, not once for the whole company. One organization can be the provider of one system and the deployer of another. The skill does not derive obligations from a lookup table. It works through them in conversation and flags where the AI Act article mapping needs your verification.

The other plugins have similar guardrails written into their instructions:

• The renewal tracker flags a cancel-by date that falls on a weekend when the contract is added to the register, so the report does not print a wrong effective deadline later.
• Tabular review requires a verbatim supporting quote and its location (section, heading or page) for each cell. A quote can only be cut at sentence boundaries, and the cut must be marked.
• The board consent drafter has a hard stop when no consents repository is configured and you have provided no seed consent. With seed consents but no repository, the draft follows house format and notes that no precedent search was available.
• NDA triage applies the governing-law and restrictive-covenant positions recorded in your commercial profile, which means your playbook choices decide the GREEN/YELLOW/RED result.

Connectors, Westlaw access and the legal-builder-hub

None of the connectors are required to start. Skills run from your practice profile and from documents you upload. Each connector you add gives a plugin read access to a specific system, and the README lists what each one covers:

• Ironclad: contract register, renewal dates and clauses (commercial-legal)
• DocuSign / DocuSign CLM: envelope status, executed contracts and CLM metadata (commercial-legal)
• iManage: matter workspaces and document versions from the DMS (commercial-legal, corporate-legal)
• Everlaw: e-discovery productions, tagged sets and chronologies (litigation-legal)
• CourtListener: federal dockets and opinions; public, with an optional API key
• Trellis: state court dockets and motions (litigation-legal)
• Definely: in-document drafting and defined-terms checks
• Descrybe: case law research and summarization
• Solve Intelligence: patent drafting and prosecution
• TopCounsel: matter routing and outside counsel panel

Aurora, Lawve AI and Courtroom5 also appear in the list. Every connector except CourtListener is marked as needing a customer subscription.

To get Westlaw access, install the separate cocounsel-legal plugin from Thomson Reuters, which sits under external_plugins. The README describes it as Westlaw Deep Research producing cited reports across caselaw, statutes, regulations, Practical Law and secondary sources. It needs your own subscription and connects through OAuth.

The README also describes a legal-builder-hub with three controls for teams that add skills from other sources. An allowlist restricts sources by default (registries, publishers, connectors, licenses). A license gate applies policy based on deployment context: personal, firm-internal or product-embedding. An install log keeps an auditable record of what was installed, from where, under what license and with what review verdict. If you plan to combine this suite with community collections such as awesome-legal-skills, the hub gives you a place to record that review.

For the Managed Agents route, the README sets credentials with export ANTHROPIC_API_KEY=sk-ant-... and runs the same system prompt and skills behind your own workflow engine. The scheduled watchers are designed to run this way, not inside an interactive Claude Code session.