The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external implementation plans from files like docs/plans/feature-plan.md, which creates a theoretical surface for indirect prompt injection.
Ingestion points: Task text is read from plan files and passed to subagents.
Boundary markers: The prompts use clear markdown headers to separate task descriptions from instructions.
Capability inventory: Subagents can write files, execute tests, and commit to git.
Sanitization: The skill does not explicitly sanitize the text of the tasks.
Mitigation: The three-stage review process (Implementer, Spec Reviewer, and Code Quality Reviewer) acts as a strong defense-in-depth mechanism, ensuring that any code produced is reviewed for both compliance and quality by independent agents.
[SAFE]: All identified patterns, including adversarial verification instructions, are legitimate development practices aimed at improving implementation accuracy and do not present a security risk.

subagent-driven-development