The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of specifications (docs/specs/<slug>.md) and user-provided plans to drive implementation tasks.
Ingestion points: SKILL.md (acceptance criteria from spec files, user plans, and analysis of external data).
Boundary markers: Lacks explicit delimiters or warnings to ignore embedded instructions within processed markdown specifications.
Capability inventory: Extensive capability surface including file system writes, network operations, and arbitrary script execution through automation tools.
Sanitization: references/boundary-discipline.md mandates strict validation and parsing at I/O boundaries for the generated code, which mitigates some downstream risks.
[DYNAMIC_EXECUTION]: The skill explicitly directs the agent to generate and run ad-hoc scripts to automate repetitive developer tasks.
Evidence: references/build-the-lever.md provides a protocol for building and executing "codemods," "generators," or "scripts" for mechanical work within the session, which may execute logic not fully reviewed by the user.

implement