The Agent Skills Directory

[SAFE]: The skill implements a robust security model for processing untrusted input. It uses a phased workflow (Analyze -> Present -> Confirm -> Implement) that prevents the AI agent from acting autonomously on external requests without human oversight.
[INDIRECT_PROMPT_INJECTION]: While the skill's primary function is to ingest potentially untrusted data (stakeholder messages from Slack, email, or tickets), the risk of indirect prompt injection is mitigated by mandatory human verification. The instructions explicitly state: 'Do not implement before user feedback' and 'Present interpretation candidates with risks', ensuring the user evaluates the technical plan before execution.
Ingestion points: External stakeholder requests provided as text arguments or context in SKILL.md.
Boundary markers: The skill employs distinct phases and a structured flow schema (.hypercore/qa/flow.json) to isolate analysis from implementation.
Capability inventory: Access to Edit/Write for codebase modifications and Bash for validation commands.
Sanitization: Human review serves as the primary validation gate for interpreting external intent.
[COMMAND_EXECUTION]: The use of shell commands is restricted to validation tasks (tests, typechecks, builds) after implementation is confirmed, which is standard behavior for development-focused skills.

qa