Blazing Rust

A Rust performance skill for Claude Code, Codex, and other agents that support Agent Skills.

The agent measures a baseline, finds the bottleneck, and pushes it toward the hardware limit one measured change at a time. It stops at lines it will not cross on its own: changed behavior, unsound unsafe, weaker security, new threads in library code, and new dependencies or build settings. Faster options beyond those lines are reported with their measured gain. It covers allocations, data structures, memory layout, async code, SIMD, parallelism, and compiler settings.

Results

Four tasks, each run with and without the skill on the same prompt, with no user instructions loaded. 27 checks cover:

• measuring before editing;
• correct results;
• not crossing the lines above;
• pushing past the first win;
• not editing code when only asked why it is slow.

What changed with the skill, on Opus:

Sonnet with the skill produced safe 12–19x speedups at about a fifth of Opus's cost. With the skill, runs take roughly twice the time and tokens of runs without it, because the agent measures more. These are single runs on an Apple M1, so treat individual numbers as rough. The test cases are in evals/.

Install

With Node.js 22.20 or later, run this from the project where you want the skill:

npx skills add painhardcore/blazing-rust --skill blazing-rust

Choose your agent when prompted, or select it with -a:

# Claude Code
npx skills add painhardcore/blazing-rust --skill blazing-rust -a claude-code

# Codex
npx skills add painhardcore/blazing-rust --skill blazing-rust -a codex

Add -g to install for all your projects. The skills CLI also supports Cursor, OpenCode, and other agents.

To install from a local checkout, replace painhardcore/blazing-rust with its absolute path.

Install without Node.js

Clone this repository, or use your existing checkout:

git clone https://github.com/painhardcore/blazing-rust.git
cd blazing-rust

Choose the destination for your agent:

For example, install for Claude Code:

skill_dir="$HOME/.claude/skills/blazing-rust"
mkdir -p "$skill_dir"
cp -i SKILL.md LICENSE "$skill_dir/"
mkdir -p "$skill_dir/references"
cp -i references/techniques.md "$skill_dir/references/"

For Codex, set skill_dir="$HOME/.agents/skills/blazing-rust" instead. For other agents, use their skills directory. Copy SKILL.md, LICENSE, and references into a folder named blazing-rust.

See the Claude Code and Codex documentation for supported skill locations.

Use

Start a new agent session after installation. Invoke the skill directly:

Claude Code:

/blazing-rust Investigate why parsing large files is slow. Measure a baseline, optimize the bottleneck, and verify the result.

Codex:

$blazing-rust Investigate why parsing large files is slow. Measure a baseline, optimize the bottleneck, and verify the result.

The agent can also load the skill when you ask about Rust performance. Give it representative inputs, a target metric, and any deployment constraints.

Steer it from the prompt

By default the agent keeps pushing until it meets your target, reaches the hardware limit, or runs out of candidates that stay within its lines. Add a sentence to your prompt to move a line:

The agent tells you the gain it left on the table, so you can rerun with a looser line if that gain is worth it.

The skill uses your project's Rust toolchain and measurement tools. It does not require Criterion or a profiler. When a tool is unavailable, the agent must state what it could not measure.

The technique guide has detailed optimization advice. The agent reads the sections relevant to the workload.

Example: CarryFS

A run on CarryFS produced these changes. The figures come from the commit notes and have not been reproduced independently.

The key change replaced object.to_be_bytes().to_vec() plus a pushed tag byte with a [u8; 9] array. Another commit cut 14,338 allocator calls per round, but its latency results were within noise, so it claims no speedup.

License

MIT. Copyright 2026 Andrei Iurchenkov.