The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as its primary function is to ingest and evaluate untrusted external skills.
Ingestion points: SKILL.md (Step 0) specifies that the agent takes a target skill path or name as input.
Boundary markers: Absent. The instructions do not define delimiters or specific safety instructions to prevent the agent from potentially obeying commands embedded within the untrusted skill being evaluated.
Capability inventory: The evaluation process involves reading file contents and analyzing metadata, which could be exploited if the agent executes instructions found in the target content.
Sanitization: Absent. No sanitization or validation of the input skill's content is described.
[EXTERNAL_DOWNLOADS]: The skill documentation and reference files mention external resources from trusted organizations and well-known services.
Evidence: References the anthropics/skills repository on GitHub and raw.githubusercontent.com for calibration and baseline benchmarking.
Evidence: References skills.sh and the vercel-labs organization for ecosystem analysis and baseline reporting.

skill-evaluation