The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted conversation history to generate documentation, creating a surface for indirect prompt injection.\n
Ingestion points: The agent reads "relevant conversation evidence" (SKILL.md) and changed source code.\n
Boundary markers: Absent; there are no instructions to use delimiters or ignore embedded instructions within the evidence data.\n
Capability inventory: The skill writes files to docs/solutions/ and executes the git CLI via the provided scripts/validate-doc-claims.py.\n
Sanitization: The skill requires strict source-grounding citations and uses regex extraction with safe (no-shell) subprocess.run calls in its validation scripts.\n- [COMMAND_EXECUTION]: The script scripts/validate-doc-claims.py executes git commands using the subprocess.run function.\n
The script passes arguments as a list, which prevents shell-based command injection vulnerabilities.\n
Execution is restricted to local repository metadata checks, such as verifying paths and commit SHAs.

capture-learning