The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests untrusted data from user requests, codebase files, and wiki pages to generate implementation plans.
Ingestion points: The skill reads AGENTS.md, codebase context via read-codebase-context, and external wiki/documentation pages to establish feature scope.
Boundary markers: Employs specific markdown fencing (four backticks) for delta.md requirement blocks to separate instructions from data.
Capability inventory: The skill is capable of writing documentation files (design.md, delta.md) and invoking downstream skills like plan-feature and implement-task.
Sanitization: The skill mitigates risks through mandatory human approval gates ("STOP", "user approves") and a formal self-review process defined in references/spec-self-review.md before execution.

brainstorm-feature