Master Claude for Legal
A community skill pack that turns Anthropic's two legal webinars into an installable playbook for running Claude correctly on privileged legal work.
SKILL.md itself shows good safety design: never auto-send, sensitive writes need approval, client data requires a commercial tier, multi-matter mixing is flagged, claims are grounded vs. flagged, and limits are stated. Deductions: a single user override ("skip the privilege confirmation") bypasses checks; the privilege-layer configuration lives in references/ files not provided for review; no rollback or isolation mechanisms are described.
As a routing meta-skill the instructions are self-consistent, the category mapping is clear, and there is a clarifying-question fallback — happy path plausible. Deductions: the key paths (15 references, 6 starter skills) were not provided in evidence, so existence and consistency cannot be verified; no error handling, failure feedback, or tests; the static cap of 10 cannot be exceeded.
Audience and scenarios are clear (firms, in-house legal, legal ops), trigger terms are specific, and non-fit boundaries (not legal advice, needs Cowork, no IT setup) are declared. Deductions: overly broad triggers ("legal work") risk false invocation; core function depends entirely on the Cowork/Anthropic commercial ecosystem, with no Chinese support — a real limitation for mainland-China users.
Layered structure with progressive disclosure, concrete install notes, thorough contribution/maintenance process, explicit MIT license, disciplined provenance. Deductions: license metadata is NOASSERTION, conflicting with the MIT LICENSE file; SKILL.md itself has no version or changelog; the promised directory contents are not all visible in this evidence, and maintenance rests on a single unverified publisher.
The meta-skill does plausibly shorten the path to the right tool and reference and adds value via framing and safe defaults. Deductions: the core value (references content, starter-skill output quality) is absent from the evidence, so direct usability of outputs is unverifiable; much value is inherent to Claude features, leaving limited evidence of independent marginal benefit; static cap of 7 applies.
Provenance discipline is strong on paper: transcripts of both webinars, the 51-question dataset, per-file Source lines in data/, and a cross-referencing blog post. Deductions: all of this is claimed rather than shown — data/ and references/ contents were not in the evidence set, so transcription quality and quote accuracy cannot be statically corroborated; multiple independently reproducible evidence types are not demonstrated.
- The privilege confirmation can be bypassed with a single user override; enterprises should enforce this via external policy, not in-skill prompts.
- Only SKILL.md was assessed; contents of references/, skills/, and data/ are unverified — check quotes and transcripts independently.
- Core function depends on Cowork and Anthropic commercial tiers with no Chinese support, limiting reachability and fit for mainland-China users.
- Repository license metadata is NOASSERTION, conflicting with the MIT LICENSE file; treat LICENSE as authoritative and request a metadata fix.
- This is a static review with no execution, low confidence; it does not constitute legal advice.
What does this skill do, and when should you use it?
Maintained by HAQQ Legal AI, this MIT-licensed pack is built from both of Anthropic's 2026 legal webinars plus HAQQ's deployment experience across roughly 9,800 firms. A SKILL.md meta-skill acts as the entry point: it classifies the user's question, loads the matching reference from fifteen documents in references/, and routes to six installable starter skills. It also ships three governance templates (firm AI policy, client data explainer, vendor security questionnaire) and the full webinar transcripts with a structured dataset of all 51 audience questions. It explicitly does not provide legal advice and cannot replace lawyers or your firm's own privilege sign-off.
The meta-skill maps each request to a category (privilege configuration, MCP hardening, connector catalog, practice-area plugins, Microsoft 365 surfaces, long documents, verification, etc.), reads the corresponding reference file before answering, then recommends the right Claude surface — Cowork, Word, Excel, PowerPoint, Outlook, or Code — and a starter skill. It enforces a four-layer privilege check before any output touching real matter data, cites sources at paragraph/page level, flags unsupported assertions, requires sensitive write actions to need approval, and never auto-sends email. The six starter skills cover NDA triage, multi-party version diff, tabular review (Excel output with cell-level citations), meeting brief, citation round-trip verification, and weekly status synthesis. references/ adds MCP permission hardening, the 20+ connector catalog, the 12 practice-area plugins, cross-surface context preservation, and a cold-start customization interview; templates/ holds the policy and questionnaire documents; data/ contains both webinar transcripts (whisper.cpp and subtitle-derived), the 51-question JSON/markdown dataset with answer status, and theme analysis.
- In-house legal teams at companies where IT enabled Cowork and connectors but legal must build its own skills and policy — start with setup-checklist.md's four-hour walkthrough
- Small and mid-size law firms piloting Claude whose case-management vendors haven't shipped MCP connectors yet — use the MCP hardening and connector catalog references
- M&A diligence teams processing many same-schema documents — install the tabular-review skill for Excel output with cell-level citations
- Lawyers sending weekly client or partner updates — install the status-synthesis skill (the Friday newsletter pattern)
- Litigators worried about hallucinated citations — run citation-verifier to round-trip every quoted source
- Researchers and journalists covering legal AI — use the data/ directory's transcripts and 51-question dataset for independent analysis
What are this skill's strengths and limitations?
- Fully traceable: every reference doc carries a Source line, every starter skill a Provenance section, and data/ preserves raw transcripts plus all 51 questions with answer status (25 answered live, 11 partial, 13 not addressed, 2 off-topic)
- Targets the gaps: directly answers the 13 questions the live session skipped, expanded with patterns from HAQQ's ~9,800-firm deployment base
- Pragmatic privilege and safety defaults: commercial-tier confirmation, approval-gated write actions, matter-scope mixing warnings, per-citation round-trip verification
- Starter skills are explicitly remixable templates, with a cold-start interview pattern to guide firm customization
- Includes a firm AI policy template, client-facing data explainer, and vendor security questionnaire that most firms need but few have written
- GitHub's license field shows NOASSERTION despite the README claiming MIT — users should verify independently
- Long-document workflows depend on Cowork or an equivalent agentic harness; chat-only Claude degrades past ~50 pages
- Only tested against Cowork, Claude Code, the Word add-in, and Team/Enterprise plans; other platforms and self-hosted MCP setups are explicitly untested
- No automated test suite; quality relies on manual maintenance
- Cannot replace legal advice — the privilege framework needs your own counsel's sign-off, and IT/RBAC configuration still requires an admin
- Content reflects Anthropic's product state as of April/May 2026 (e.g., cross-surface context preservation) and may age as the platform evolves
How do you install this skill?
Three options. Cowork (recommended): git clone https://github.com/sboghossian/master-claude-for-legal.git, then in Cowork go to Customize → Plugins → Local → Add and point to the cloned directory. Claude Code: clone into ~/.claude/plugins/master-claude-for-legal, then run /plugin install master-claude-for-legal in a session. Other Anthropic-format tools: drop the skills/ directory into your tool's skills path. No installation steps are documented for Claude.ai web or bare API environments.
How do you use this skill?
For first-time users, follow the four-hour quickstart: 1) read references/setup-checklist.md; 2) install Anthropic's legal plugin (Cowork → Customize → Plugins → Anthropic and Partners → Legal); 3) pick one skill from skills/ matching a weekly workflow; 4) remix it using references/skill-authoring.md. The meta-skill auto-invokes when the user mentions legal work, contracts, redlines, NDAs, privilege, court filings, depositions, regulatory compliance, M&A diligence, or Claude in Word/Excel/PowerPoint/Outlook; you can also name a reference explicitly (e.g., 'use the privilege-layers reference'). Long-document tasks (50+ pages) require Cowork or an equivalent agentic harness — chat-only Claude hits context rot.
How does this skill compare with similar options?
It composes with — rather than replaces — Anthropic's official legal plugin: the pack assumes that plugin is installed, and its six starter skills extend the official NDA triage, contract review, and meeting brief capabilities. When a firm has its own remixed versions, the meta-skill defers to the firm version.