The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as specifications and repository files to generate its output plans.
Ingestion points: Instructions in SKILL.md for analyzing project specifications and existing codebases to derive tasks.
Boundary markers: The instructions include an explicit requirement to treat repository evidence as untrusted data and to ignore any instructions embedded within them. The generated output uses a structured plan header to isolate global constraints.
Capability inventory: The skill generates markdown documentation in docs/plans/ and suggests standard development commands (git, pytest) for task verification.
Sanitization: Relies on explicit behavioral instructions to ignore embedded commands and integrates with external validation tools like cross-review for plan verification.

plan-crafting