improve
An agent skill that audits a codebase and writes self-contained implementation plans for cheaper agents to execute.
At a Glance
MIT-licensed agent skill that audits a codebase and writes implementation plans; free to install and use.
Engagement
Available On
Listed Nov 2026
About improve
improve is an open-source agent skill from shadcn that audits a codebase and writes implementation plans in markdown. It is designed so a capable model handles analysis and spec-writing while cheaper models or humans carry out the changes. The skill itself never modifies source code; its output is the plan files.
What It Is
improve is a skill in the Agent Skills format, installed with npx skills add shadcn/improve. Run it with /improve inside a compatible agent to map the repository, audit it, and produce prioritized findings. Selected findings become one markdown plan each in a plans/ directory, along with an index containing recommended order and a dependency graph.
How the Audit Works
The workflow has five stages: recon, audit, vet, prioritize, and plan. Recon identifies the stack, conventions, and the repo's build, test, and lint commands, and it also reads ADRs, PRDs, and similar intent docs when present. The audit fans out parallel subagents across nine categories: correctness, security, performance, test coverage, tech debt, dependencies and migrations, DX, docs, and direction. Every finding carries file:line evidence, impact, effort, and confidence. The advisor re-reads each cited location to drop false positives and records rejections so they do not return. Findings are then ordered by leverage.
Executable Plans and Closing the Loop
Plans are written for a weak executor: they inline file paths, code excerpts, and conventions, end each step with a verification command, and list out-of-scope items and STOP conditions. Each plan stamps the git commit it was written against for a drift check. Commands include quick, deep, focused (security, perf, tests, bugs), and branch-scoped audits, plus plan, review-plan, execute, and reconcile. execute dispatches a cheaper subagent in an isolated git worktree and reviews its diff against the plan, leaving merging to the user. --issues publishes plans as GitHub issues.
Safety Rules
The README states the skill only writes to plans/, does not run commands that mutate the working tree, never reproduces secret values, and declines direct implementation requests.
Community Discussions
Be the first to start a conversation about improve
Share your experience with improve, ask questions, or help others learn from your insights.
Pricing
Open Source
MIT-licensed agent skill that audits a codebase and writes implementation plans; free to install and use.
- MIT License
- Install via npx skills add shadcn/improve
- Works in any agent supporting Agent Skills format
- Plans written as plain markdown
- Optional publishing of plans as GitHub issues (--issues)
Capabilities
Key Features
- Full, quick, deep, focused, and branch-scoped codebase audits
- Findings with file:line evidence, impact, effort, and confidence
- Vetting step that drops false positives and records rejections
- Self-contained markdown plans with verification gates and STOP conditions
- Feature suggestions grounded in repo evidence via /improve next
- Plan review, execution via cheaper subagent in isolated git worktree, and backlog reconcile
- Optional publishing of plans as GitHub issues
- Reads ADRs, PRDs, and design docs to respect decided tradeoffs
