Agent Skills

A collection of portable skills for AI agents.

The performance-engineering skill helps an agent diagnose bottlenecks, form testable hypotheses, prioritize high-leverage changes, design representative benchmarks, and keep only verified improvements. It is inspired primarily by the engineering principles in the Abseil Performance Guide and converts those principles into a portable investigation workflow rather than copying the source material.

This is an independent project and is not affiliated with or endorsed by Google or the Abseil project.

Compatibility

The skill follows the open Agent Skills specification. It does not depend on a specific model, vendor, coding agent, profiler, benchmark library, or programming language.

It can be installed by clients supported by the skills CLI, including Codex, Claude Code, Cursor, GitHub Copilot, Gemini CLI, OpenCode, Windsurf, and others.

Repository structure

skills/performance-engineering/
├── SKILL.md
└── references/
    ├── benchmarking.md
    ├── concurrency-and-io.md
    ├── estimation.md
    ├── language-notes.md
    ├── measurement.md
    ├── mechanisms.md
    └── sources.md

Repository-level evaluation scenarios live under evals/; they are not part of the installed skill.

Install locally

From this repository:

npx skills add . --skill performance-engineering

List the skills discoverable in the repository without installing:

npx skills add . --list

Install from GitHub

From GitHub:

npx skills add sasicodes/skills --skill performance-engineering

Choose a target agent interactively, or pass the CLI's --agent option.

Example requests

• “Why did p99 latency regress after this change?”
• “Review this hot loop for meaningful performance problems.”
• “Create a representative benchmark for these two implementations.”
• “Reduce allocation and GC pressure in this service.”
• “Profile this application and identify the actual bottleneck.”
• “This microbenchmark improved, but production became slower. Investigate why.”

Design principles

• Diagnose before optimizing.
• Prefer measurements over intuition.
• Eliminate work and fix algorithms before micro-optimizing instructions.
• Connect every proposal to a mechanism, tradeoff, and verification plan.
• Adapt advice to the language, runtime, workload, and deployment environment.
• Preserve correctness, security, and maintainability.
• Revert changes that do not produce a meaningful verified improvement.

Validation

Validate the skill's format with the Agent Skills reference implementation:

skills-ref validate skills/performance-engineering

Also verify discovery with:

npx skills add . --list

The files under evals/ specify manual comparison scenarios and grading criteria. They are development evidence, not executable tests, and are not installed with the skill.

License

MIT. See LICENSE.