Beautiful Rust

Three standard Agent Skills and a lints config that teach coding agents to write, review, port, and enforce idiomatic Rust:

• rust-best-practices loads whenever an agent writes or changes Rust, and steers it toward idiomatic, robust, maintainable code.
• rust-review reviews existing Rust code, a file, a crate, or a diff, and writes a prioritized report where every finding cites the rule it breaks and shows the fix.
• rust-port ports C, C++, Python, Go, or JavaScript to idiomatic Rust with the same behavior, porting the original tests first so you can prove the two match.
• lints/ is a drop-in clippy configuration that enforces the machine-checkable rules in CI.

They work with Claude Code, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, and any other agent that reads the Agent Skills format. Every code example in the skills compiles on stable Rust, and every lint in the config was checked against current clippy.

The guidance is distilled from five sources, with a condensed digest of each bundled in the skill so the agent can consult the original reasoning when it needs to:

• Google Rust Style Guide
• Rust API Guidelines (official rust-lang checklist)
• Effective Rust by David Drysdale
• Rust Design Patterns (idioms, patterns, anti-patterns)
• Microsoft Pragmatic Rust Guidelines

Where the sources disagree, the skill picks a side and says why.

What it changes

Asked for a small semver-parsing library, an agent without the skill produced working code with public fields, no # Errors docs, and unwrap() in its doc examples. With the skill, the same agent kept the fields private, documented failure modes, and wrote doc examples users can copy safely. Condensed from those two runs:

// Without the skill
/// A semantic version. The fields are public for convenience.
pub struct Version {
    pub major: u64,
    pub minor: u64,
    pub patch: u64,
    pub pre: Vec<Identifier>,
    pub build: Vec<String>,
}

impl Version {
    /// Parses a version string.
    ///
    /// ```
    /// let v = Version::parse("1.2.3").unwrap();
    /// ```
    pub fn parse(s: &str) -> Result<Version, ParseError> { /* ... */ }
}

// With the skill
/// A semantic version. Equality, hashing, and ordering ignore build metadata.
pub struct Version {
    major: u64,
    minor: u64,
    patch: u64,
    pre: Vec<Identifier>,
    build: Vec<String>,
}

impl Version {
    /// Parses a version string according to SemVer 2.0.0.
    ///
    /// # Errors
    ///
    /// Returns [`ParseVersionError`] if the string is empty, a numeric part has a
    /// leading zero or overflows `u64`, or an identifier contains invalid characters.
    ///
    /// ```
    /// # use semver_lite::Version;
    /// # fn main() -> Result<(), semver_lite::ParseVersionError> {
    /// let v = Version::parse("1.2.3-beta.1")?;
    /// assert_eq!(v.major(), 1);
    /// # Ok(())
    /// # }
    /// ```
    pub fn parse(s: &str) -> Result<Self, ParseVersionError> { /* ... */ }

    /// Returns the major version number.
    pub fn major(&self) -> u64 { self.major }
}

Across three test tasks (a library, a CLI, and refactoring a messy module), the skill passed 29 of 30 checks against 24 of 30 without it. See Testing the skills.

Install

Each skill is a folder under skills/: rust-best-practices/, rust-review/, and rust-port/. Every option below puts those folders where your agent looks for skills. Install all three; the review and port skills lean on the best-practices skill for style.

Any agent, one command

The skills installer supports most agents (Node 22.20 or later):

npx skills add morganlinton/BeautifulRust

Add -g to install for your user instead of the current project, or -a <agent> to pick an agent, such as -a codex or -a cursor.

Claude Code

Install it as a plugin so updates arrive with one command:

claude plugin install beautiful-rust --marketplace morganlinton/BeautifulRust

Inside a session, the same thing is:

/plugin marketplace add morganlinton/BeautifulRust
/plugin install beautiful-rust@beautiful-rust

This installs all three skills. They appear as /beautiful-rust:rust-best-practices, /beautiful-rust:rust-review, and /beautiful-rust:rust-port, though you rarely need to call them, since they trigger on their own. To try it for one session without installing, clone the repo and run claude --plugin-dir ./BeautifulRust.

GitHub Copilot

With GitHub CLI 2.90 or later:

gh skill install morganlinton/BeautifulRust rust-best-practices
gh skill install morganlinton/BeautifulRust rust-review
gh skill install morganlinton/BeautifulRust rust-port

This installs into the current project by default. Add --scope user to install for your user, or --agent <name> to install for a different agent.

Gemini CLI

gemini skills install https://github.com/morganlinton/BeautifulRust --path skills/rust-best-practices
gemini skills install https://github.com/morganlinton/BeautifulRust --path skills/rust-review
gemini skills install https://github.com/morganlinton/BeautifulRust --path skills/rust-port

OpenAI Codex, Cursor, or anything else: copy the folder

Clone the repo and copy the three skill folders into your agent's skills directory. Most agents now read the shared .agents/skills/ location:

git clone https://github.com/morganlinton/BeautifulRust
mkdir -p ~/.agents/skills
cp -R BeautifulRust/skills/rust-* ~/.agents/skills/

Agent	User directory	Project directory
Claude Code	~/.claude/skills/	.claude/skills/
OpenAI Codex	~/.agents/skills/	.agents/skills/
GitHub Copilot	~/.copilot/skills/ or ~/.agents/skills/	.github/skills/ or .agents/skills/
Cursor	~/.cursor/skills/ or ~/.agents/skills/	.cursor/skills/ or .agents/skills/
Gemini CLI	~/.gemini/skills/ or ~/.agents/skills/	.gemini/skills/ or .agents/skills/

Claude Code does not read .agents/skills/, so use its own directory or the plugin. Restart the agent if the skill doesn't show up right away.

claude.ai

claude.ai takes one zip per skill, with the skill folder as the zip root:

cd skills
zip -r ../rust-best-practices.zip rust-best-practices -x '*.DS_Store' -x '*/evals/*'
zip -r ../rust-review.zip rust-review -x '*.DS_Store' -x '*/evals/*'
zip -r ../rust-port.zip rust-port -x '*.DS_Store' -x '*/evals/*'

Upload each under Customize > Skills. Code execution must be enabled first.

Reviewing existing code

Ask your agent to review a file, crate, or diff, for example "review src/parser.rs" or "what would you change in this PR". The rust-review skill runs cargo fmt, clippy, and the tests when a toolchain is available, then writes a report with three tiers: must fix, should fix, and consider. Every finding gives a file:line location, the rule it breaks with its source (an API Guidelines code such as C-GETTER, an Effective Rust item, a Microsoft guideline, or a clippy lint), and the smallest code change that fixes it. Each report ends with what the code already does well.

examples/review-inventory.md is a real review the skill produced for the deliberately messy module in the eval cases.

Porting code to Rust

Ask your agent to port a module, for example "port parser.c to Rust" or "rewrite this Python service in Rust". The rust-port skill reads the original and its tests, ports the tests first, designs a Rust API instead of transliterating the original, and verifies with cargo test and clippy. It maps each language's habits to their Rust equivalents: NULL and None to Option, error codes and exceptions to an error enum, out-parameters to return values, flag ints and magic strings to typed enums and options, and dict insertion order to an ordered collection instead of a HashMap that would silently reorder output. Each port comes with a PORTING.md that maps the old API and tests to the new ones and lists every deliberate difference.

Bevy

Bevy calls your systems and builds their arguments, so some general Rust advice is wrong for them. Taken at face value, the skills would tell an agent to change Res<Time> to a reference, replace use bevy::prelude::* with explicit imports, and split a system to quiet clippy. When a crate depends on bevy or bevy_ecs, both the writing and review skills now read references/bevy.md first. It covers Bevy 0.20 and does two things:

• It names the conventions to leave alone: system parameters taken by value, the Bevy prelude, public fields on plain-data components, and complex query types.
• It lists the ECS bugs that compile cleanly. Two queries that can both mutate the same component panic with error B0001 on the first frame. A system can read the world before another system's Commands have applied. Systems that depend on each other may have no order between them. A .single().unwrap() crashes the game once the player dies.

Every behavior in the reference was checked with a test against bevy 0.20.0. Two eval cases cover it: a review of a combat plugin with two of those bugs planted, and a feature added to the fixed plugin, which checks that the agent keeps Bevy's conventions.

Enforcing the rules with clippy

The skills teach the rules; the lints keep them from slipping back in. Copy lints/clippy.toml next to your Cargo.toml and paste the [lints] block from lints/lints.toml into it, then run:

cargo clippy --all-targets --all-features -- -D warnings

It flags unwrap() outside tests, unsafe blocks without a // SAFETY: comment, #[allow] without a reason, any bool function parameter, silent lossy as casts, missing # Errors and # Panics docs, and leftover dbg! and todo!. Lints marked [lib] in the file are for libraries, and applications can drop them. Bevy crates need four extra allows, listed in the Bevy reference, because the table otherwise flags every idiomatic system signature.

Every lint in the table is proven to fire. lints/check/ is a small crate with one deliberate violation per lint, and this script runs clippy on it from the repo root:

bash scripts/check-lints.sh

It fails if a lint stops firing, if something unexpected fires, or if the check crate's copy of the table drifts from lints/lints.toml. allow_attributes and allow_attributes_without_reason need rust-version = "1.81" or later, because clippy switches them off for crates that can't use #[expect].

What the writing skill covers

skills/rust-best-practices/SKILL.md is about 500 lines, organized so an agent can scan it while coding:

1. Mindset: Rust-shaped code, invalid states unrepresentable, application vs library
2. Pre-handback checklist: fmt, clippy with -D warnings, test, doc; lint policy
3. Ownership and borrowing: what to do instead of cloning for the borrow checker
4. Types and API design: naming, newtypes, common traits, builders, visibility, generics vs dyn
5. Errors and panics: Option vs Result vs panic, library vs application error types
6. Everyday idioms: iterators, combinators, format!, imports, drop guards
7. Concurrency and async: Send assertions, lock discipline, async fn, yield points
8. Unsafe and FFI: when it's allowed, SAFETY: comments, soundness, bindgen
9. Crates, modules, dependencies: workspace layout, features, semver
10. Documentation and comments
11. Tests
12. Performance, in order
13. Final checklist

skills/rust-best-practices/references/ holds one digest per source for deeper lookups.

Layout

BeautifulRust/
├── .claude-plugin/
│   ├── plugin.json         plugin manifest
│   └── marketplace.json    lets this repo act as its own marketplace
├── skills/
│   ├── rust-best-practices/   writing skill
│   │   ├── SKILL.md
│   │   ├── references/        one digest per source, plus bevy.md
│   │   └── evals/
│   ├── rust-review/           review skill
│   │   ├── SKILL.md
│   │   └── evals/
│   └── rust-port/             porting skill
│       ├── SKILL.md
│       └── evals/             a C module and a Python module to port
├── lints/
│   ├── lints.toml             [lints] block for Cargo.toml
│   ├── clippy.toml            clippy thresholds
│   └── check/                 one deliberate violation per lint
├── scripts/
│   ├── check-skill.py         validates skills, evals, and manifests
│   ├── check-lints.sh         proves every lint in the table fires
│   ├── check-examples.py      compiles every Rust example in the docs
│   └── examples.json          context that lets fragment examples compile
├── benchmarks/
│   └── cross-agent-2026-10/   Claude Code vs Codex, with and without the skills
├── examples/
│   └── review-inventory.md    a real review report
├── CHANGELOG.md
├── CONTRIBUTING.md
└── README.md

Testing the skills

skills/rust-best-practices/evals/ holds four test cases in the layout claude plugin eval reads: a library API design task, an application CLI task, and a refactor-this-module task with an input file. Each case has the prompt, a description of a good result, and one assertion per file under graders/. Run them with and without the skill from the skill directory:

claude plugin eval . --allow-tools Write,Edit,Bash --scaffold

skills/rust-review/evals/ holds a review case for the same messy module and one for a Bevy plugin with two ECS bugs planted, and skills/rust-port/evals/ holds a C module and a Python module to port, each with its original test suite. The writing skill also has a Bevy case that adds a feature to a working plugin. Run them the same way from each skill's directory.

These are real model runs and are billed.

Across agents. The October 2026 benchmark ran all eight cases on Claude Code and OpenAI's Codex CLI, with and without the skills, and graded the results blind:

	No skills	Skills
Claude Code (Opus 5.5)	72/85 (85%)	80/85 (94%)
Codex CLI (default model)	53/85 (62%)	76/85 (89%)

The skills helped Codex more than Claude, and helped reviews most. They hurt in one case: Claude did worse on the Bevy feature task with them. The benchmark page has the details, the method, and the caveats. Notably, it was one run per configuration, graded by a Claude model.

Updating

Edit the skill, bump version in .claude-plugin/plugin.json, commit, and push. Claude Code plugin users pick up the new version with:

claude plugin update beautiful-rust

Anyone who copied the folder or used an installer re-runs the install command. Record each release in CHANGELOG.md.

Validate before pushing. python3 scripts/check-skill.py checks skill frontmatter, links, eval cases, and plugin manifests. Pass --json for machine-readable output. python3 scripts/check-examples.py compiles every Rust block in the skills, reference digests, README, and examples with cargo check.

python3 scripts/check-skill.py
python3 scripts/check-examples.py
claude plugin validate .

Contributing

Issues and pull requests are welcome. Every rule needs a published source; see CONTRIBUTING.md for how to propose a rule, report a review false positive, or add an eval case.

Sources and attribution

The guidance in this skill is a synthesis. The five digests in skills/rust-best-practices/references/ are paraphrased summaries written for this project; they restate each source's advice in condensed form and link back to the original. Read the originals for the full reasoning.

Source	Author	License
Google Rust Style Guide	Google	CC BY 3.0
Rust API Guidelines	The Rust Project	MIT or Apache-2.0
Effective Rust	David Drysdale, published by O'Reilly	CC BY-NC-ND 4.0 (online edition)
Rust Design Patterns	rust-unofficial contributors	MPL-2.0
Pragmatic Rust Guidelines	Microsoft	MIT

The skill text, the digests, the banner, and the generator script in this repository are MIT licensed (see LICENSE). The crab is a crab, not the Rust mascot.

License

MIT. See LICENSE.