improve is an agent skill released by shadcn: after running /improve, it first identifies the repository's tech stack and build, test, and lint commands, then lists an audit of issues for you to select which items to address. For each selected item, it writes a self-contained implementation plan under plans/, including excerpts of existing code, specific steps, verification commands, and conditions to stop and report if reality does not match the plan. The skill itself does not modify code; you can also use /improve execute to delegate implementation to a cheaper model in an isolated worktree, then review changes against the plan, with humans deciding on merges.
Why it matters Expensive models are used only for judgment and writing specifications, while execution is delegated to smaller models; plans include verification commands and conditions to stop when unexpected issues arise.
Stars 9711 (+49 / 24h)
Reasoning
- 49 new stars in 24 hours (indicating attention only)
- Supports scoping to the current branch only, focusing solely on security or performance, and publishing plans as GitHub issues
Usage
- Install with npx skills add shadcn/improve, then run /improve in the repository (or /improve quick to save usage)
Risks
- Full audits consume significant strong-model usage; plan quality depends on the completeness of repository documentation and tests