The Agent Skills Directory

[SAFE]: The skill performs read-only audits of project metadata (manifests, task runners, CI configurations) using standard command-line tools like git, grep, and awk to gather context.
[SAFE]: The file-writing process in agents-finalize.sh uses an atomic write pattern (writing to a temporary file then moving it) and enforces strict size and quality limits to prevent resource exhaustion or poor model performance.
[SAFE]: The skill includes a 'filler' check that prevents the agent from generating generic or redundant instructions, which acts as a quality control mechanism.
[INDIRECT_PROMPT_INJECTION]: The skill reads external data from the repository (README, existing AGENTS.md, docs) which could theoretically contain malicious instructions. However, the risk is minimized by the skill's internal logic which filters output through a strict guide and a linting script that flags generic or verbatim duplicated content. The overall severity remains low as this is a standard capability for development-oriented agents.
[DYNAMIC_EXECUTION]: The audit script uses node -e or python3 -c to reliably parse JSON manifests (like package.json) when available. This is a standard and safe implementation for local metadata extraction.

pflow-agents