The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, which could contain adversarial instructions intended to influence the sharpening process or inject malicious rules into agent instructions.
Ingestion points: Untrusted data enters the agent context through incident reports (Step 1), global memory files (~/.claude/memory/), and repository-level learnings or retrospects (Batch Mode).
Boundary markers: The skill employs clear instructions to "distill principles" and "strip specifics." It implements a "prohibited-subject gate" that requires checking terms against a denylist before drafting edits.
Capability inventory: The skill possesses significant capabilities, including the ability to execute shell commands (Bash), modify instruction files (Edit), write new files (Write), and spawn recursive agent sessions (Agent).
Sanitization: Automated sanitization is performed via a mandatory "Mechanical overfit scan" covering nine categories (person names, ticket IDs, repo names, etc.) and the execution of pre-commit hooks (scrub-staged.sh) that scan for identifiers and prohibited terms.
[COMMAND_EXECUTION]: The skill frequently utilizes the Bash tool to interact with the version control system and execute local validation scripts.
Evidence: Step 5 and Step 8 of SKILL.md involve running repository hooks (.githooks/check-*.sh, .githooks/scrub-staged.sh) and standard git operations (git add, git commit, git push, git diff). These commands are used for maintaining the integrity and versioning of the skill files.

wk-sharpen