The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on bundled Node.js scripts to perform repository-wide audits and backups of Claude context files. These scripts are implemented with significant attention to local security, including the use of O_NOFOLLOW flags and filesystem identity snapshots to prevent symlink race conditions and path traversal attacks. Additionally, the backup script verifies that storage locations remain outside Git worktrees to prevent accidental exposure of sensitive local state.
[PROMPT_INJECTION]: The skill processes untrusted user-controlled content from .claude/rules, CLAUDE.md, and auto-memory files, which creates an indirect prompt injection surface where embedded instructions could attempt to influence the agent's auditing logic.
Ingestion points: Data enters the context via the inventory-claude-memory.mjs and scan-claude-memory-risks.mjs scripts, which read files from the repository root, .claude/ directory, and the user's home directory.
Boundary markers: The skill instructs the agent to use the assets/review-report-template.md to isolate findings and maintains a clear separation between read-only audit phases and mutation phases.
Capability inventory: The agent can execute the provided Node.js scripts, create report files, and perform atomic edits to markdown context files if authorized by a user-approved cleanup plan.
Sanitization: The scan-claude-memory-risks.mjs script proactively redacts secrets, tokens, and PII from its output and specifically flags "absolute language" (e.g., 'always', 'must', 'forbidden') to help the agent identify and neutralize potentially malicious instructions embedded in the audited data.

claude-memory-curator