The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions involve executing local Python scripts (harness/scripts/failure-flywheel.py) to manage failure logs and generate reports. These scripts are internal components of the skill's architecture and perform tasks consistent with the skill's documented purpose.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data (failure summaries and details) provided by the user or agent, which is then processed to generate draft rule stubs. This presents an indirect prompt injection surface.
Ingestion points: Failure descriptions and details provided as arguments to the record command in SKILL.md.
Boundary markers: The skill uses shell quoting for the record command arguments and structured JSONL for storage, but specific prompt delimiters for the final interpolated draft content are not detailed.
Capability inventory: The skill can execute local Python scripts and write draft files to the llmwiki/wiki/sources/draft/ directory.
Sanitization: There is no explicit documentation regarding the sanitization of failure summaries; however, the risk is mitigated by the mandatory human review step (W04) which requires a person to approve and finalize any generated draft before it becomes an active rule.

failure-flywheel