The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill evaluates external, potentially attacker-controlled data sources like usage logs, project tickets, and git history to justify recommendations for code or feature removal. A malicious actor could inject instructions into these data sources to bias the agent's assessment or trick it into proposing the removal of important (but non-security-critical) components.
Ingestion points: Usage logs, git history, tickets, surveys, and stakeholder confirmation, as defined in SKILL.md (Trigger Guidance) and reference/evaluation-criteria.md.
Boundary markers: The skill instructions in SKILL.md include explicit "Never" rules regarding safety-critical code (auth, encryption, input validation) and "Ask First" gates for public APIs or data-integrity targets.
Capability inventory: The skill is strictly advisory and read-only. SKILL.md explicitly forbids the agent from editing code, documents, or executing deletion work directly, routing these tasks to other agents like Sweep or Zen.
Sanitization: No specific data sanitization is mentioned; the skill relies on evidence-based quantification (Cost-of-Keeping score) and human-in-the-loop routing for its final recommendations.

void