The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests and processes historical conversation logs which may contain untrusted data or previous malicious attempts. \n
Ingestion points: Reads JSONL files from sensitive directories ~/.claude/projects/ and ~/.codex/sessions/ via scripts/extract-conversations.py. \n
Boundary markers: The extraction script adds structured headers like [USER]:, [ASSISTANT]:, [PROVIDER]:, and [SOURCE_FILE]: to the processed text, providing context to analysis agents. \n
Capability inventory: The skill can spawn sub-agents (TeamCreate), execute shell scripts, run Python code, and write updates to brain files and other skills' SKILL.md files (Step 6). \n
Sanitization: The Python script performs basic cleaning (whitespace, system reminder filtering) and truncates message length, but does not perform instruction filtering or safety checks on the message content. \n- [COMMAND_EXECUTION]: The skill executes multiple commands to process data. \n
Evidence: SKILL.md invokes a snapshot script (sh .claude/skills/meditate/scripts/snapshot.sh) and the local extraction script (python3 "$SKILL_DIR/scripts/extract-conversations.py"). It also performs directory cleanup via rm -rf and creates temporary files in /tmp/. These operations are consistent with the documented process.

ruminate