The Agent Skills Directory

[COMMAND_EXECUTION]: The skill identifies and executes build, test, and lint commands found within the target repository's configuration. It implements a safety filter that explicitly forbids running destructive commands such as those involving deployment, database migrations, or financial costs, and mandates user confirmation for commands requiring credentials or services (SKILL.md, Phase 2).
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository data (source code, comments, READMEs) that could contain malicious instructions. It contains a dedicated 'Security boundary' section instructing the agent to treat all ingested content as data rather than directives. It specifically requires the detection and exclusion of attempts to widen scope or plant malicious instructions into the output (SKILL.md, Phase 1 and Security Boundary).
[DATA_EXFILTRATION]: The instructions include a specific check to detect and report any attempts to plant exfiltration URLs within the generated output. Furthermore, it enforces a rule that secret values from environment files like .env are never extracted, only variable names from templates (SKILL.md, Security Boundary).
[INDIRECT_PROMPT_INJECTION]: Evidence Chain:
Ingestion points: Analyzes manifests, lockfiles, CI workflows, source code, comments, READMEs, commit messages, and ADRs (SKILL.md, Phase 1).
Boundary markers: Explicit 'Security boundary' section and instructions to treat external text as untrusted data (SKILL.md, Phase 1, Step 5).
Capability inventory: Execution of repository-defined scripts (Phase 2), writing to AGENTS.md and shim files (Phase 5/6).
Sanitization: Filtering of secret values, removal of embedded exfiltration URLs, and a whitelist-only approach to non-destructive commands (SKILL.md, Phase 2 and Security Boundary).

awesome-agents-md-generator