The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local shell script scripts/snapshot.sh to recursively find and aggregate markdown files from the project structure into temporary files in /tmp/.
[COMMAND_EXECUTION]: The 'Apply changes' step (Step 6) grants the agent the authority to perform destructive operations on the filesystem, including deleting 'low-value' or 'redundant' notes and rewriting the core project instruction file (CLAUDE.md).
[DATA_EXFILTRATION]: The skill accesses and reads highly sensitive agent state from the internal auto-memory directory located at ~/.claude/projects/<project>/memory/.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted markdown content from the 'brain' vault which then influences subagent decisions regarding which files to delete or modify.
Ingestion points: Reads all files in brain/ and .agents/skills/ via the snapshotting script.
Boundary markers: Files in the snapshot are delimited with path headers, but the subagent prompts in references/agents.md lack explicit instructions to disregard any embedded commands or instructions found within the audited notes.
Capability inventory: The skill has broad capabilities to delete files, merge content, and rewrite critical project configuration files (CLAUDE.md).
Sanitization: No sanitization or safety filtering is performed on the content of the notes before they are passed to the subagents for analysis.

meditate