Reference · plugin fact sheet
ai-governance-legal
The ground-truth summary behind Module 6. Every line below was read from the plugin's
actual files (version 1.0.2, checked 2026-07-17) — paths cite the
live repo.
(Accuracy note: the plugin's own README lists only 8 of the 10 skills — it omits
ai-inventory and customize — and calls policy-monitor an “agent”;
it is a skill. The table below reflects what's on disk.)
Identity
From .claude-plugin/plugin.json:
“Triages proposed AI use cases against your registry, runs impact assessments across the regimes in
scope, reviews vendor AI terms for training-on-data and liability gaps, and keeps your AI policy current
with practice.”
The 10 skills, by job
| Job | Skills |
|---|---|
| Set up & tune the profile | cold-start-interview (regimes in scope; refuses a company-level AI role when the EU is in footprint — roles are per-system; shadow-AI discovery writes unapproved tools into the registry as [UNDOCUMENTED — NEEDS TRIAGE]), customize (one change at a time; refuses to drop load-bearing verification tags), matter-workspace (multi-client only — off for in-house) |
| The two registries — the front door | use-case-triage (APPROVED / CONDITIONAL / NOT APPROVED against the use-case registry + red lines; no-match → CONDITIONAL pending an AIA; batch mode; a “provisional” escape on an unconfigured profile), ai-inventory (ai-systems.yaml: one record per AI system with role, tier, and basis — Article 5 prohibited screen → Annex III high-risk → GPAI → limited/minimal; never classifies silently) |
| The deep paperwork | aia-generation (AI impact assessment per regime in scope; an AIA is not a PIA; provider-vs-deployer split table when the company wears both hats; risk-quality bar — “2-5 real risks, not 12 padded ones”), vendor-ai-review (13-term review of vendor AI terms — training-on-data is “the one most people miss”; the AI-addendum gap check; stacked-vendor/flow-down test when your vendor wraps someone else's model) |
| Keeping current — two directions of change | reg-gap-analysis (outside-in: a new AI law vs. your posture — scope first, prohibited-practice gaps outrank everything, “accepted gaps” documented as first-class output), policy-monitor (inside-out: practice vs. your written policy — sweep or direct-query; the canonical catch: a policy that says “we do not use AI in employment decisions” after an AIA approved exactly that) |
| Day zero | policy-starter (first-cut AI policy sourced from published model policies — NIST, bar guidance, peers — never invented; scope interview first; every judgment call flagged [review]) |
No subagents, no hooks — and no research connector
Nothing scheduled runs here: no agents/, no hooks. The “monitor” is a skill you run.
From .mcp.json:
Slack and Google Drive only — no research/statute tool,
in the practice area that moves fastest. The plugin compensates with graduated source tags
([settled] / [verify] / [verify-pinpoint] — pinpoints are the
highest fabrication risk: “EU AI Act article numbers in particular shifted during consolidation”), the
no-silent-supplement stop (research thin → the skill asks you which source to accept), and
references/currency-watch.md —
which distrusts itself: “If the last-verified date above is more than 90 days old, treat this file as
stale.”
State it keeps on your machine
~/.claude/plugins/config/claude-for-legal/ai-governance-legal/CLAUDE.md— the practice profile. The use-case registry, red lines, governance tiers, and vendor-AI positions live inside it — this file IS the institutional memory the triage reads and writes.…/ai-systems.yaml— the per-system AI inventory (role, tier, basis, EU nexus, next review). Deliberately does not auto-derive obligations: “a hardcoded role × tier → obligations table is exactly the kind of confident-and-wrong artifact that ends up in a board memo.”…/verification-log.md— one line per verified cite/fact, so the next reviewer doesn't re-verify.~/.claude/plugins/config/claude-for-legal/company-profile.md— shared across all 12 plugins.
Conventions worth knowing
- The registry is authoritative. “Generic AI ethics reasoning is not a substitute for what this company has actually decided.” No match ≠ no — it's CONDITIONAL, and the decision writes back.
- Red lines are absolute — in both directions. “Do not soften red line outcomes. If it's a no, it's a no.” And for non-lawyers the plugin gates NOT APPROVED as heavily as APPROVED: “wrongly rejecting a use case is also a consequential error.”
- Roles are per-system, not per-company. One organization can be provider of System A and deployer of System B; substantially modifying a vendor's model can make you the provider of the modified system.
- Severity floors carry. The triage's risk tier flows into the AIA as a floor — no silent demotion downstream.
- Retroactive triage. “Already piloting” is a documented edge case: the review gap itself goes in the record — “a gap to document, not to wave through.”
Sources: the plugin's CLAUDE.md, .mcp.json,
references/currency-watch.md, and the skill files — all under
ai-governance-legal/ on GitHub.