The Agent Skills Directory

[SAFE]: The skill architecture demonstrates defensive design, utilizing a critic agent to validate all outputs against an 8-point quality gate, which includes checks for root-cause anchoring and anti-generic tests. Support scripts for environment setup include protections against common attacks like symlink traversal.- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as external content from web searches and local files are ingested and processed. 1. Ingestion points: research-agent.md fetches external data via WebSearch/WebFetch; SKILL.md reads local artifacts like diagnose.md and product-context.md. 2. Boundary markers: Absent in agent prompts. 3. Capability inventory: SKILL.md (Bash tool), research-agent.md (WebSearch and WebFetch tools), scripts/bootstrap-experience.ts (file system writes). 4. Sanitization: No explicit sanitization or filtering of external content before interpolation into prompts. The multi-agent sequential review process acts as a structural mitigation against simple injection attempts.

prioritize