The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes local project scripts such as scripts/perf/perf_grid_matrix.sh and scripts/perf/compare_perf_runs.py, alongside standard git commands (git diff, git checkout, git worktree), to manage performance data and identify regressions. These operations are correctly scoped to the local repository and the project's performance data directory.
[DATA_EXFILTRATION]: Includes instructions to perform a git pull from the primary project repository to capture an up-to-date baseline. This network operation is restricted to the authoritative project source and does not involve exfiltration of sensitive information or communication with unknown external domains.
[PROMPT_INJECTION]: Instructions are strictly focused on technical performance analysis and optimization workflows. There are no attempts to override safety guardrails, use deceptive roleplay, or extract system-level prompts.
[SAFE]: The 'Self-improve' mechanism encourages the agent to document new findings in local reference files and extract repeatable logic into local scripts. This practice promotes maintainability and automation within the project environment without introducing external dependencies or risky behaviors.

optimize