The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from Claude Code session logs (.jsonl) and Git history. These logs contain full conversation history, including user inputs and past assistant responses, which may contain malicious instructions designed to influence future behavior.
Ingestion points: ~/.claude/projects/{project-dir}/{session-id}.jsonl (logs), git log, and git diff output.
Boundary markers: The instructions do not define strict boundary markers or "ignore instructions" directives when processing the logs or diffs.
Capability inventory: The skill has the capability to suggest and apply changes to core configuration files like CLAUDE.md, SKILL.md, and agent definitions using the Edit tool based on its analysis (as described in SKILL.md Step 7).
Sanitization: There is no evidence of sanitizing or filtering instructions found within the logs or commit history before they are processed and used to generate configuration improvement suggestions.\n- [COMMAND_EXECUTION]: The skill executes local shell scripts and git commands to extract and analyze data.
Evidence: SKILL.md and README.md describe executing scripts/analyze-session-logs.sh, as well as scripts from sibling skills like extracting-session-data to parse session metadata and metrics.

retrospecting