The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses highly sensitive session logs and transcripts from several agent platforms (Claude Code, Codex, Cursor, and Antigravity) stored in the user's home directory (e.g., ~/.claude, ~/.cursor). These files contain a comprehensive history of user prompts, agent reasoning, and tool execution logs.
[DATA_EXFILTRATION]: The helper script collect-messages.py performs a broad scan of the filesystem starting from the root directory (/) to resolve Cursor project slugs into absolute paths, which involves reading directory names across the system.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from past session transcripts to generate improvements for future agent behavior.
Ingestion points: Session log files (.jsonl) and database files (.db) located in ~/.claude, ~/.codex, ~/.cursor, and ~/.gemini as processed by the collect-messages.py script.
Boundary markers: The skill instructions specify filtering for human-typed messages and paraphrasing findings rather than quoting raw transcript text, though the effectiveness of these boundaries depends on the model's adherence during the mining process.
Capability inventory: The skill has the capability to modify project instruction files (like AGENTS.md and CLAUDE.md) and create or edit other agent skills using the create-agent-skill and writing-for-agents tools.
Sanitization: While the collection script filters for human-typed messages, the subsequent 'mining' step involves interpreting transcripts which may contain unsanitized reasoning or tool results from previous sessions that could contain malicious instructions.

retrospect