The Agent Skills Directory

[PROMPT_INJECTION]: The skill documentation in SKILL.md contains strings typical of prompt injection attacks, such as "ignore previous rules" and "skip validation". These are used exclusively as examples of malicious patterns the agent should recognize and document as untrusted data rather than follow as instructions. This is a defensive pattern intended to improve agent robustness.
[COMMAND_EXECUTION]: The skill workflow involves executing standard local development commands including npm test, npm run lint, and npm run test:coverage. To mitigate risks from external inputs, the skill mandates a safety checklist requiring human approval for any commands derived from external plan files.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting and acting upon content from *.plan.md files.
Ingestion points: Reading implementation plans from user-provided or session-persistent *.plan.md files.
Boundary markers: Instructions explicitly state that plan content must be treated as data, not as instructions to the AI.
Capability inventory: The agent is authorized to run test commands, write code files, and perform Git commits.
Sanitization: The skill includes a "Plan safety checklist" that requires human review for shell commands and explicitly rejects destructive filesystem operations.

tdd-workflow