Development Skills

A collection of Agent Skills for Rust and frontend development. Compatible with Claude Code, OpenAI Codex, Cursor, and other agent platforms.

Overview

Skill	Purpose
code-review-rust	Review Rust for idioms, security, complexity, duplication, test quality, and NATS patterns. One finding per file in .backlog/tasks/. Also usable as an implementation guardrail.
code-review-web	Review React/TypeScript (Vite) frontends for hooks, types, async, a11y, security, and tests. Same finding and guardrail pattern as the Rust skill.
code-review-triage	Group triaged backlog findings into semantic review waves (code-review-plan-waveN parents).
code-review-run-wave	Run one planned wave in an isolated git worktree: apply fixes, QA, merge, close.
code-review-run-waves	Run every open wave concurrently (one worktree each); land merges one at a time via a shared lock.
commit-script	Analyze git state and generate a script that stages grouped files into conventional commits — optionally on a topic branch that ends in a gh pull request.
rust-make-clippy-pedantic	Lint a clean Rust checkout at pedantic strength via flags only, file one pedantic-labelled backlog task per warning — test-only style findings, generated files and out-of-tree warnings dropped, high-volume lints aggregated per crate — estimate the cleanup, and show (or with --apply, write) the matching Cargo.toml / clippy.toml lint policy.
rust-meta	Process external Rust content and integrate new knowledge into code-review-rust.

Installation

Option 1: Claude Code Plugin Marketplace

The repository is a plugin marketplace with a single dev-skills plugin containing every skill — they reference each other's files (review waves lean on the worktree protocol, rust-meta on the review rule set), so they ship together.

claude plugin marketplace add rsvalerio/ai
claude plugin install dev-skills@rsvalerio

Inside a session, /plugin marketplace add rsvalerio/ai and /plugin install dev-skills@rsvalerio work the same way. Skills are invoked through the plugin namespace — /dev-skills:code-review-rust — and track the latest commit on main; update with claude plugin update dev-skills@rsvalerio.

Option 2: Clone the Repository

git clone https://github.com/rsvalerio/ai.git
cd ai

# Claude Code
cp -r skills/* .claude/skills/

# OpenAI Codex
cp -r skills/* .codex/skills/

For a live symlink into ~/.claude/skills/ while developing this repo, use make link (see AGENTS.md).

Option 3: Install Individual Skills

Using agent-skills-cli:

agent-skills install https://github.com/rsvalerio/ai/tree/main/skills/code-review-rust

Restart your AI tool after installing so it picks up the new skills.

Requirements

• AI agent: Claude Code, OpenAI Codex, Cursor, or another Agent Skills-compatible platform
• Developing this repo: Git, plus rumdl and skill-validator at the versions pinned in .tool-versions — mise install gets both, or make install-tools via Homebrew — and claude-code on PATH for make validate-marketplace (the asdf route is in CONTRIBUTING.md). Full workflow in AGENTS.md

Usage

Skills load from context or explicit invocation.

code-review-rust

• "Run a code-review-rust on this crate."
• "Review this code for anti-patterns, ownership issues, and unsafe usage."
• "Check this module for security vulnerabilities and OWASP violations."
• "Analyze complexity and readability of this file."
• "Find duplicated logic between these two modules."
• "Review my tests for effectiveness, gaps, and flakiness."
• "Audit NATS consumer patterns against best practices."

Parallel instances are fine — each finding is its own file under .backlog/tasks/.

code-review-web

• "Run a code-review-web on this frontend."
• "Review these React components for hooks and accessibility issues."
• "Check this Vite app for XSS and Web Crypto misuse."

Implementation guardrail mode

code-review-rust (and code-review-web) can load rules before edits so violations do not ship. Description-matching alone is best-effort; for deterministic activation, add a Cursor rule, CLAUDE.md directive, or AGENTS.md directive.

Full snippets, vendoring when the skill is missing, and verification: docs/implementation-guardrail.md.

CLI agents

claude --model opus -p "/code-review-rust @crates"
opencode run "/code-review-rust @crates"

Review waves

code-review-triage groups findings into waves; code-review-run-wave / code-review-run-waves execute them. Each wave gets its own git worktree; merges serialize through a lock. A failed merge parks the worktree/branch so work stays resumable.

claude -p "/code-review-triage"       # finish triage first (single-writer)
claude -p "/code-review-run-waves"    # all open waves
claude -p "/code-review-run-wave"     # one wave

Details: Worktree Protocol.

rust-make-clippy-pedantic

• "Run rust-make-clippy-pedantic on this workspace."
• "How much work is it to get this crate clippy-pedantic clean?"
• "Configure this workspace for pedantic clippy — show me the config first."

Requires a clean git status; the run aborts rather than stashing or pulling. Lint levels are configured with -W flags after -- rather than by editing the crate. What keeps the run off your files is the rest of it: --locked so Cargo cannot write Cargo.lock, a scratch CARGO_TARGET_DIR so target/ is untouched, and no --fix. Findings are written to .backlog/, which is the point.

The run finishes by printing the Cargo.toml lint tables and clippy.toml that would make the strictness permanent: [workspace.lints.*] plus a [lints] workspace = true opt-in per member for a workspace, or direct [lints.rust] / [lints.clippy] tables for a single crate. Pass --apply to have it write them — the only mode that edits checked-in lint configuration, though every run writes its findings to .backlog/. Neither commits.

rust-meta

• "Process this external Rust doc and integrate new knowledge into the rules."
• "Evaluate this blog post against our evaluation criteria."

Resources

• Agent Skills Specification
• Agent Skills Integration Guide
• Agent Skills Marketplace
• openai/skills Repository
• skill-validator

Contributing

Setup, gates, commit format, and pull request rules: CONTRIBUTING.md. Skill authoring conventions and publishing: AGENTS.md.

main is protected: pull requests merge by squash only, need signed commits, and must pass the Lint, Validate and Install checks. make ci runs the same gates locally.

License

Apache-2.0. See LICENSE.

Support

• Issues: https://github.com/rsvalerio/ai/issues

Acknowledgments

Rust rules draw on the Rust API Guidelines, The Rust Book and Rust By Example. Frontend rules draw on the React documentation, the TypeScript handbook and the OWASP Top 10, alongside community best practices.