The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted documentation files from the project repository to extract tasks and update project documentation.
Ingestion points: Project files like README.md, FEATURES.md, and historical status reports are ingested to generate TODO_LIST.md and ROADMAP.md entries.
Boundary markers: No explicit delimiters or boundary instructions are implemented to prevent the agent from following malicious instructions potentially embedded in the processed documentation.
Capability inventory: The skill relies on the agent's ability to write to files and execute shell commands for build verification.
Sanitization: Edits use regex for structural matching, but the skill lacks semantic sanitization of the processed text content.
[DYNAMIC_EXECUTION]: Helper scripts included in the skill assets use dynamic module loading for shared logic.
Evidence: The scripts annotate-prose.py and annotate-rows.py use importlib.util.spec_from_file_location to load the annotate-markers.py module from a path constructed at runtime relative to the script location.
[COMMAND_EXECUTION]: The verification process instructs the agent to execute shell commands defined within the project being audited.
Evidence: SKILL.md directs the agent to "Detect the build system (nix flake check, cargo test, pnpm test, scripts/check-skills.sh) and run the canonical command — do not substitute," which involves executing potentially untrusted project-specific scripts.

docs-health