The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses scripts/agents-context.sh to audit the repository by executing standard utilities such as git, find, grep, awk, and sed. These operations are used to identify the project stack, task runners, and existing documentation to inform the generation of agent instructions.
[DYNAMIC_EXECUTION]: The json_keys function in the context script uses brief Node.js and Python commands (node -e and python3 -c) to parse keys from local manifest files like package.json and composer.json. This is limited to extracting script names for the audit and does not execute arbitrary code from these files.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it reads project documentation (README.md, etc.) and interpolation this content into the agent's drafting process.
Ingestion points: Reads docs/, README.md, CONTRIBUTING.md, and manifest files.
Boundary markers: The agents-md-guide.md provides strict rules (Removal, Derivability, and Duplication tests) to prevent the agent from blindly following or copying external text.
Capability inventory: The skill can write to AGENTS.md and CLAUDE.md via the agents-finalize.sh script.
Sanitization: The agents-finalize.sh script acts as a linter, checking for verbatim duplication of documentation, filler phrases, and placeholders, which serves to mitigate the risk of importing malicious instructions from project files.

pflow-agents