improve: A code audit skill that plans with expensive models and executes with cheap ones

shadcn/improve is an agent skill that uses the most powerful LLMs to analyze a codebase and produce an actionable implementation plan. This tool does not directly perform code modifications; it handles only the stages where intelligence is concentrated, such as analysis and specification definition. The final output is a plan in the form of a Markdown file that other agents or developers can execute immediately.

It audits 9 categories, including accuracy, security, performance, and test coverage, using parallel subagents. Each finding is presented with the file path, line number, impact, effort, and confidence, and is prioritized based on leverage. If documents such as ADRs or PRDs exist, it references them to filter out already-decided trade-offs to avoid duplicate reporting.

The generated plan has a self-contained structure with fully embedded context. It specifies current code snippets, verification commands, and stop conditions on failure, enabling smaller models to perform tasks without misunderstanding. The /improve execute command runs a cheaper model in an isolated worktree, reviews the output like a tech lead, and then decides whether to approve or reject it.

The core principle is generating only plan files without modifying source code. It performs read-only analysis, and actual code modifications are made in an isolated environment after user approval. It is suitable for developers who want to systematically improve a complex codebase but find the cost of deploying expensive models to every line burdensome.