The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/diff_surgeon.py and the pre-commit hook hooks/karpathy-gate.sh utilize subprocess.run and shell commands to interact with git. These operations are used to retrieve diffs and run analysis tools on the local repository. The implementation in diff_surgeon.py uses a list-based argument structure without shell=True, which is a safe practice to prevent command injection.
[INDIRECT_PROMPT_INJECTION]: Several scripts (assumption_linter.py, complexity_checker.py, goal_verifier.py) ingest and analyze external text such as project plans and source code. While these are ingestion points for potentially untrusted data, the tools perform heuristic regex-based analysis or code complexity metrics and do not execute the content as code, significantly limiting the risk of injection affecting the agent's logic. The agent instructions in agents/karpathy-reviewer.md direct the agent to review these outputs, which is a standard and low-risk use case.
[SAFE]: The skill is a legitimate productivity toolset. It uses only Python standard libraries, provides clear documentation and expected output examples, and its behavior aligns perfectly with its stated purpose of enforcing coding discipline.

karpathy-coder