The Agent Skills Directory

[SAFE]: The skill provides a process-oriented playbook for performance tuning. It contains no executable scripts, hardcoded credentials, or external network triggers. It relies on the agent's built-in capabilities to manage git operations and subagents within the host environment.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface common to development tools, as it involves reading project source code and processing user-provided metrics to generate code modifications. This risk is a byproduct of its primary function. The skill includes built-in mitigations by mandating supervisor review of all code diffs before they are committed.
Ingestion points: Project source code and workload data analyzed to reproduce performance symptoms (SKILL.md, Step 1).
Boundary markers: Not explicitly defined for ingested code, but the workflow uses a 'frozen' measurement harness to provide objective validation.
Capability inventory: File system modification via git commands, subagent delegation for code generation, and invocation of the 'benchmark-checklist' tool (SKILL.md, Steps 2, 5, and 8).
Sanitization: Risk is mitigated through mandatory human/supervisor verification of all subagent outputs and the requirement for a deterministic verification script (references/prove-it-works.md).

hillclimb