The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input in the form of git diffs and pull request metadata to perform its security analysis.
Ingestion points: The skill reads external code changes via git diff and user-provided descriptions in Step 2.
Boundary markers: The methodology requires explicit context research (Phase 1) and comparative analysis (Phase 2) to establish boundaries between existing code and new changes.
Capability inventory: The skill utilizes the Bash, Read, Grep, and Glob tools to inspect the filesystem and execute local scripts.
Sanitization: The instruction set includes a mandatory "Confirm each candidate yourself" step (Step 4) which requires the agent to manually trace data flows and verify exploitability, serving as a defensive layer against deceptive content within the analyzed code.
[COMMAND_EXECUTION]: The skill executes a local Bash script to calculate the diff scope and determine which files to audit.
Evidence: SKILL.md invokes scripts/review-scope.sh with user-supplied arguments captured from $ARGUMENTS.
Mitigation: The review-scope.sh script employs a strict character whitelist ([A-Za-z0-9_./=:~@+,%\ '"-]) and rejects any input containing shell-significant characters (like semicolons or backticks) before processing them, effectively mitigating command injection risks.

review-security