Skip to content

SKILL.md

Latest commit

 

History

History
67 lines (51 loc) · 7.21 KB

SKILL.md

File metadata and controls

67 lines (51 loc) · 7.21 KB
Raw
Error in user YAML: (<unknown>): mapping values are not allowed in this context at line 2 column 339
---
name: rust-coding-skill
description: Use whenever writing, reviewing, refactoring, or generating Rust — any .rs file, Cargo project, or request like "write a Rust function/CLI/library", "fix this Rust", "make this idiomatic", "satisfy the borrow checker", or "clean up clippy warnings". This skill makes code-writing agents produce clean, correct, idiomatic Rust: it enforces type-system-first data modeling so invalid states are unrepresentable, Result/? error handling instead of panics, iterator chains over index loops, ownership and borrowing that the borrow checker accepts without needless clones or unwraps, and a mandatory cargo fmt + cargo clippy -D warnings + cargo test gate before any work is called done. Covers application, library, systems, and network/security tooling code alike. Apply it proactively on every Rust task, not only when asked.
---

Idiomatic Rust

When to use

Use for any task that produces or changes Rust: new code, refactors, reviews, bug fixes, or generating snippets. If a .rs file, Cargo.toml, or Rust idiom is in scope, this skill governs how you work. Default to it; do not wait to be asked.

Core operating procedure

Follow this loop in order every time you write Rust. Each step prevents a class of defect the next step cannot fix.

  1. Model the data with the type system first. Before writing logic, design the types. Use enums for "one of" and structs for "all of". Make invalid states unrepresentable — if a value can never legally be two things at once, do not store two fields; use an enum. Wrap primitives in newtypes (struct UserId(u64)) when they carry meaning, so the compiler rejects mixing them. Derive Debug, and Clone/PartialEq/Eq/Hash/Default where they make sense. See references/data-modeling.md.
  2. Choose ownership deliberately. Decide what owns each value and what merely borrows it. Take &str/&[T] parameters rather than &String/&Vec<T>; return owned values rather than dangling references; accept impl Into<String> or generic borrows where it improves the API. Reach for clone() only when a borrow genuinely cannot satisfy the lifetimes — never to silence the borrow checker. See references/ownership-borrowing.md.
  3. Handle errors with Result and ?, not panics. Library code returns Result<T, E> / Option<T>; it does not unwrap, expect, or panic! on conditions a caller could hit. Define a custom error type (thiserror for libraries, anyhow for applications) and let ? propagate with From conversions. Reserve panics for true bugs (broken invariants), and prefer expect("invariant: ...") over unwrap() when you must. See references/error-handling.md.
  4. Prefer iterators and combinators over manual loops and indexing. Express transformations as lazy adapter chains (iter().filter().map().collect()); use collect::<Result<_, _>>() to short-circuit fallible work; use the entry API for map updates. Index into slices only when the algorithm truly needs positions. See references/collections-iterators.md.
  5. Use traits, generics, and lifetimes idiomatically. Bound generics with the traits they need; prefer static dispatch, reaching for dyn only for heterogeneous collections or to shrink code size; implement From/TryFrom for conversions; lean on elision and add explicit lifetimes only when required. See references/traits-generics-lifetimes.md.
  6. Run cargo fmt, then cargo clippy -- -D warnings, and fix every finding. Do not suppress lints with #[allow(...)] unless you can justify it in a comment. Clippy's suggestions are idiom guidance — apply them, do not silence them.
  7. Write tests. Add #[cfg(test)] unit tests for logic and edge cases, integration tests for public behavior, and doc tests for examples. Cover the error paths, not just the happy path. See references/testing.md.

Non-negotiables

  • No needless clone() — clone only when ownership is genuinely required, never to dodge the borrow checker.
  • No unwrap() / expect() / panic! in library code on caller-reachable conditions; propagate with ?. Panics are for bugs only.
  • Never swallow errors (let _ = result;, empty match arms, .ok() to discard). Propagate or handle them meaningfully.
  • No unsafe without a // SAFETY: comment documenting the invariant that makes it sound, and the encapsulation that upholds it.
  • Make invalid states unrepresentable; do not validate at runtime what the type system can forbid at compile time.
  • Code must pass cargo clippy -- -D warnings with zero #[allow] escape hatches that lack a written justification.
  • Prefer borrowing over owning in signatures (&str, &[T], impl AsRef<_>); return owned data, never references to locals.
  • Follow standard naming: snake_case items, CamelCase types, SCREAMING_SNAKE_CASE consts, new/with_*/try_* constructors.

Reference map

Open the relevant file on demand — read it when the task touches its area, not upfront.

Task / trigger Reference
Moves, borrows, lifetimes, clone avoidance, slices, thread closures references/ownership-borrowing.md
Result/Option, ?, custom errors, thiserror/anyhow, panics, CLI exit codes references/error-handling.md
Enums/structs, newtypes, typestate, making invalid states unrepresentable, builders, derives, Default references/data-modeling.md
Generics, trait bounds, associated types, static vs dyn dispatch, impl Trait, conversions references/traits-generics-lifetimes.md
Iteration, collect, entry API, Vec capacity, UTF-8-safe String handling references/collections-iterators.md
Cargo, dependency versions, modules, visibility, re-exports, semver, features, workspaces, CI references/project-structure.md
Box/Rc/Arc/Weak, RefCell/Cell, Deref, Drop/RAII references/smart-pointers.md
Threads, channels, Arc<Mutex<_>>, atomics/ordering, thread pools, Send/Sync references/concurrency.md
Test placement, assertions, panic/Result tests, integration/doc tests, TDD, fuzzing references/testing.md
Naming/casing, constructors, receiver choice, borrowed args, type aliases, operator overloading references/api-design-naming.md
unsafe contracts, raw pointers, transmute, repr, FFI, calling/exposing C, no_std references/unsafe-ffi.md
Bindings/mutability, expression style, control flow, closures, imports, conversions, macros, ranges references/idioms-antipatterns.md
Network/security tooling: async scanners, bounded concurrency, timeouts/rate limits, untrusted-input parsing, secrets handling, TLS, static binaries, CLIs references/security-tooling.md

Verification

You are not done until all three pass cleanly. Run them and fix every finding before claiming completion — do not report success on unverified code.

cargo fmt --check
cargo clippy --all-targets -- -D warnings
cargo test

Or run the bundled gate: scripts/check.sh (POSIX) / scripts/check.ps1 (Windows) — it runs all three and stops on the first failure.

If formatting differs, run cargo fmt and re-check. If clippy reports lints, fix the code (do not blanket-allow). If tests fail, fix the implementation or the test, then re-run the full gate. Iterate until all three are green.