The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from the local repository and external audit/RFC files, creating an attack surface where embedded instructions could influence the agent's behavior.
Ingestion points: The skill reads docs/arc/audits/*-audit.md, docs/arc/plans/*-refactor-rfc.md, and any user-specified code paths during its intake and hotspot scan phases (SKILL.md Step 1 and 2).
Boundary markers: The skill uses a <boundary> tag to define its scope and explicitly references references/subagent-safety.md, which contains instructions to treat repository content as data rather than instructions.
Capability inventory: The skill can dispatch subagents (Explore agents), invoke other skills (detail skill), and write new files (docs/arc/plans/ implementation plans and INDEX.md).
Sanitization: The process includes a vetting step (Step 3) where findings are cross-referenced against current code excerpts, providing a manual/logic check on the data being processed.
[COMMAND_EXECUTION]: The skill executes local Python scripts to map the codebase and identify 'god files'.
Evidence: python3 scripts/codebase-map.py . --format markdown and python3 scripts/find-god-files.py . --max-files 40 are called during the Hotspot scan in SKILL.md. These scripts are part of the Arc bundle's internal paths.

improve