Skip to content

_POLICY.md

Latest commit

 

History

History
45 lines (33 loc) · 1.97 KB

_POLICY.md

File metadata and controls

45 lines (33 loc) · 1.97 KB

Skills Policy

Risk Tiers

All skills in this catalog are assigned a risk tier. The tier determines what a skill is allowed to do and how it can be invoked.

Tier 0 — Safe (instruction-only)

  • No scripts or executable code.
  • Read-only guidance, checklists, templates, and review workflows.
  • May be model-invocable (agent can activate without user prompt).
  • No network access required.

Tier 1 — Low side-effects

  • Scripts allowed only if idempotent and local (no network).
  • Prefer tool restrictions (allowed-tools) where supported.
  • Generally model-invocable, but restrict tools defensively.

Tier 2 — High side-effects

  • Anything that pushes, deploys, deletes, or modifies external systems.
  • Remote MCP server usage (data crosses a network boundary).
  • Must be user-invoked only (never model-invocable).
  • Require explicit rollback plan documented in the skill.
  • Tool restrictions mandatory where supported.

Security Rules

  • Treat every skill bundle like a dependency: review, pin, and audit.
  • Default skills to instruction-only; scripts require explicit justification.
  • No downloading or executing remote code from within skills.
  • No curl | bash patterns or equivalent.
  • No network access unless explicitly required and declared.
  • Never run Tier 2 skills without explicit user intent.

Design Rules

  • Keep SKILL.md under 500 lines. Move detail into references/ and assets/.
  • Write descriptions like a router: specific enough to trigger correctly, narrow enough to avoid false activation.
  • Version skills via version in frontmatter.
  • Include license in frontmatter (required for cross-tool compatibility).
  • Required frontmatter fields: name, description, version, tier, license.
  • Skills do not replace always-on instructions in AGENTS.md or AGENT_INSTRUCTIONS.md.
  • Prefer internal over external (see AGENT_INSTRUCTIONS.md § Norms). Do not introduce external dependencies when an existing skill or native capability covers the need.