Performance Hints Agent Skills

A small set of skills that gives coding agents a direct route into Performance Hints, the performance guide by Jeff Dean and Sanjay Ghemawat.

_{Screenshot of the canonical source page, captured August 26, 2026. Click the image to read the article.}

The article covers many different kinds of performance work. This repository maps its 15 main guidance sections to 15 focused skills, with one additional router that selects the useful sections for the task at hand.

The article is not copied into the skills. Each focused skill fetches its section from Abseil when the agent needs it, so the source stays with its authors and updates remain available from the canonical page.

Why this exists

In Jeff Dean's Startup School 2026 interview, he discussed a performance-optimization skill that he and Sanjay Ghemawat had written internally. He then pointed to their public Performance Hints guide.

This project is an independent agent interface for that public guide. It is not their internal skill, and it is not affiliated with or endorsed by Jeff Dean, Sanjay Ghemawat, Google, or Abseil.

How it works

flowchart TB
    subgraph route["1. Route"]
        direction LR
        A["Performance question"] --> B["performance-hints<br/>router"] --> C["Focused skill or skills"]
    end

    subgraph retrieve["2. Retrieve"]
        direction LR
        D["fetch_section.py<br/>with an h2 section ID"] --> E["Download the live<br/>Abseil page"] --> F["Slice the selected h2<br/>up to the next h2"]
    end

    subgraph use["3. Use"]
        direction LR
        G{"Output mode"} --> H["HTML<br/>source fragment"]
        G --> I["Text<br/>markup removed"]
        H --> J["Agent applies the guidance<br/>to the code and evidence"]
        I --> J
    end

    route --> retrieve --> use

Loading

The retrieval path is intentionally simple:

1. The router looks at the performance question and chooses one or more focused skills.
2. Each focused skill points to a canonical <h2 id="..."> on the source page.
3. The fetcher downloads https://abseil.io/fast/hints at runtime.
4. It returns everything from that <h2> through, but not including, the next <h2>. Nested headings, examples, and code remain with their parent section.
5. The agent uses that source alongside the actual code, profile, benchmark, or cost estimate.

HTML mode returns the live source fragment, aside from trimming surrounding whitespace. Text mode removes HTML tags, decodes entities, and cleans up repeated spaces and blank lines. It does not summarize or rewrite the authors' prose. An agent may still paraphrase when explaining or applying the guidance.

No article sections are cached or committed here. The only source-page material included in the repository is the attributed header screenshot above. This keeps the skills small and the guidance current, but it also means retrieval requires network access and depends on the page retaining its <h2 id="..."> structure.

Skill map

Skill	Canonical source section
performance-hints	Router for all sections below
performance-thinking	The importance of thinking about performance
performance-estimation	Estimation
performance-measurement	Measurement
performance-api-design	API considerations
performance-algorithms	Algorithmic improvements
performance-memory-representation	Better memory representation
performance-allocations	Reduce allocations
performance-unnecessary-work	Avoid unnecessary work
performance-code-size	Code size considerations
performance-parallelism	Parallelization and synchronization
performance-protobuf	Protocol Buffer advice
performance-cpp	C++-specific advice
performance-bulk-operations	Bulk operations
performance-case-studies	CLs that demonstrate multiple techniques
performance-further-reading	Further reading

Install

Clone the repository, then copy the skill directories into the location your agent scans. For Codex:

git clone https://github.com/boshjerns/performance-hints-agent-skills.git
mkdir -p ~/.codex/skills
cp -R performance-hints-agent-skills/skills/performance-* ~/.codex/skills/

Open a new agent session so the skills are discovered.

Use it

Start with the router when a problem crosses multiple areas:

Use $performance-hints to review this request path for evidence-backed performance improvements. Start with the profile and route to only the relevant sections.

Call a focused skill when the problem is already clear:

Use $performance-parallelism to review this worker for lock contention.

Use $performance-bulk-operations to see whether batching can remove repeated boundary overhead.

You can also read a section directly from the repository:

# Readable text. The prose is not summarized or rewritten.
python3 skills/performance-hints/scripts/fetch_section.py measurement --output text

# Live source HTML for the selected section.
python3 skills/performance-hints/scripts/fetch_section.py measurement --output html

# Every h2 section ID currently present on the source page.
python3 skills/performance-hints/scripts/fetch_section.py --list

Source and license

All performance guidance accessed through these skills comes from:

Jeffrey Dean & Sanjay Ghemawat, Performance Hints, 2025, https://abseil.io/fast/hints.html

The routing metadata and retrieval code in this repository are available under the MIT License. The license applies only to this repository's original work. It does not relicense the Performance Hints article, the Abseil website, or the source-page screenshot included for identification and attribution.