The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies and processes untrusted external data, including code diffs, commit messages, and planning documents like plan.md and phase-*.md.
Ingestion points: Code repositories accessed via git commands and markdown planning files ingested from the project directory.
Boundary markers: The protocol uses structured markdown headers and adversarial analysis prompts to frame the input, although it does not use explicit safety delimiters to wrap untrusted data.
Capability inventory: The skill allows for shell command execution (git, node), sub-agent spawning (spawn_agent), and user interaction via questioning tools.
Sanitization: The 'SKEPTIC' mindset and 'Anti-Bias Gate' serve as cognitive sanitization, requiring the agent to challenge the input's logic rather than execute embedded commands.
[COMMAND_EXECUTION]: The skill executes local shell commands to retrieve data and manage review metadata.
Evidence: Employs git show and git diff to establish the scope of code changes for review.
Evidence: Executes a local repository script via node .claude/hooks/lib/review-receipt.cjs to generate snapshots and receipts of the review process.

why-review