Callstack hands your AI agent its React Native playbook

What it does

It’s a library of Agent Skills — structured markdown documents an AI coding assistant loads on demand — covering React Native performance optimization, GitHub workflow patterns, GitHub Actions build artifacts, React Native upgrades, and brownfield migration into existing native apps. The performance material is distilled from Callstack’s Ultimate Guide to React Native Optimization, spanning JS profiling and re-renders, native iOS/Android tooling, and bundle-size work like tree shaking and R8. Each skill has a SKILL.md entry point plus a references/ tree of topic files, so the agent reads the summary first and drills in only when needed.

The interesting bit

The real engineering here is packaging, not prose: the repo ships the same skills in every dialect the assistant ecosystem expects — Claude marketplace metadata, Codex plugin bundles, Cursor .mdc rules, Gemini and OpenCode skill paths — so one source of truth works everywhere. It’s glue code, but the genuinely annoying kind of glue, done thoroughly.

Key highlights

• Five skills: React Native best practices, GitHub workflows, GitHub Actions for RN builds, RN upgrades, and brownfield migration via @callstack/react-native-brownfield
• Skills follow the agentskills.io spec and are written to be actionable, searchable, and complete with pitfalls
• Runnable code examples (React Compiler, R8, dedicated RN SDKs) live in a companion repo, callstack/optimization-best-practices
• Works with assistants that have no native skills support — you just point the model at the relevant file

Caveats

• Visual profiler work (flame graphs, treemaps, Instruments) still often needs exported reports or a human eyeball; the roadmap admits agents can’t fully inspect those GUIs yet
• Device-driven verification depends on an agent-device tool that may not exist in your environment
• Debugging, testing, and CI/CD skills are listed as future work, not shipped

Verdict

If an AI assistant touches your React Native codebase, this is a cheap way to make it less confidently wrong about profiling and upgrades. Pure-web or non-RN teams can skip it — though the GitHub workflow skills are mildly useful anywhere.