AI skill pack

llms.txt gives a coding agent the facts about the SDK. The skill pack gives it the procedures — the closed loops that compose the CLI, the provider, and the build-time scanner into repeatable workflows. It ships as a Claude Code plugin, versioned in the monorepo and generated from the same frozen SDK types as llms.txt, so a skill can never reference a jurisdiction or diagnostic code that no longer exists.

Policy generates policy documents; it is not legal advice. Have a lawyer review your policies before publication.

Install#

In Claude Code:

/plugin marketplace add jamiedavenport/policystack
/plugin install policystack

The four skills then activate automatically when a task matches.

The skills#

• policystack-init — scaffold Policy in a project: run @policystack/cli init, then wire the single <PolicyStack> provider (it supplies both the policy context and the consent store from one config).
• policystack-audit — the closed loop: run policy validate --json, explain each issue code against the frozen 1.0 diagnostic surface, propose a minimal config fix, and re-validate until the config is clean.
• policystack-jurisdiction — explain the consent-model and policy-text posture implied by a declared jurisdictions set, read straight from the canonical jurisdiction table.
• policystack-instrument — find un-annotated data collection and data egress in a codebase and add collecting() / sharing() / thirdParty() / defineCookie() call sites so the generated policy matches reality.

Why it can't drift#

Every enumeration a skill cites — jurisdiction ids, lawful bases, the issue codes validate() emits — is rendered from the live @policystack/core / @policystack/sdk tables at generation time and snapshotted with a drift test that fails the build if the generated pack and the shipped files disagree. The same mechanism backs llms.txt. Removing or renaming a frozen code is a loud test failure, not a silently stale skill.

You can also point any other agent at the local reference: policy init writes policystack.llms.txt into your project — see the CLI page.