The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted historical agent transcripts, creating a vulnerability surface where malicious instructions embedded in the logs could attempt to influence the audit result.
Ingestion points: The skill mines data from agent transcripts and logs located in ~/.claude/projects/ and ~/.codex/sessions/.
Boundary markers: SKILL.md explicitly mandates that models ignore any directive-shaped content inside transcripts and treat the corpus text only as evidence for categorization.
Capability inventory: The skill allows the model to identify correction events and propose instruction changes; however, it requires manual human approval for every line proposal and provides citations for verification.
Sanitization: The skill includes a dedicated scripts/friction-scan.py script that implements a redact function using a comprehensive regex to mask API keys, tokens, and private keys before any data is processed.

history-audit