What is improve?
improve is an agent skill from shadcn, the author of shadcn/ui. You run `/improve` in a repository and the skill maps the stack and its build, test and lint commands, audits the code across nine categories (correctness, security, performance, tests, tech debt, dependencies, developer experience, docs and product direction), and returns a table of findings. Each finding cites a file and line, with impact, effort and confidence. You choose which ones become plans.
The plans are the product. Each is a Markdown file in `plans/` written for a model that has never seen the audit: current code pasted in, exact steps, a command and expected result after every step, an out-of-scope list and STOP conditions for when the code does not match. The skill never edits source itself; `/improve execute` hands a plan to a separate executor in a git worktree and reviews the diff. The repository has 12,083 stars, an MIT licence, 23 commits and no tagged releases; the last commit landed on 12 September 2026.
Is improve worth it?
The idea behind improve is an economic one: spend the strong model on judgement and the cheap model on typing. In our reading of the skill, the discipline that makes this work is in the plan template. Every step ends with a verification command and its expected output, the plan records the git commit it was written against so the executor can check for drift, and STOP conditions tell a small model to halt and report rather than improvise.
The audit is built to resist the usual failure of AI code review, which is confident nonsense. Subagents over-report, the skill says, so the advisor reopens every cited location before showing a finding, drops false positives, corrects wrong attributions and records rejections so they are not raised again next run. The README's example run on shadcn/ui shows both sides: an O(n²) loop at a named line kept, a proxy-variable "SSRF" thrown out as by design.
Depth is a cost dial. `quick` uses at most one subagent and reports about 6 high-confidence findings; the default allows up to 4 concurrent subagents; `deep` allows up to 8, one per category, across every package. When `execute` runs, the executor defaults to Sonnet unless you name another model, and a plan gets at most 2 rounds of revision before it is blocked.
The catch: no token counts or savings are published, so the claim that cheap execution pays for expensive auditing is untested in the repository. A deep run on a large codebase means several long subagent sessions on your strongest model. The `--issues` flag posts plans as GitHub issues; the skill checks visibility and asks before publishing security findings on a public repository, but a confirmation prompt is a thin barrier against posting a credential's location to the world. There are no releases, and nothing has been committed since September.
Our take: improve is most useful as a habit rather than a one-off, a monthly `/improve quick` and a `/improve branch` before each pull request, with the plans read by a person before any model acts on them.
Where it earns its place
A tech lead inheriting a mid-sized TypeScript service
Find the few fixes worth doing first, with evidence
Run `/improve` and reply with the findings to plan, such as "plan 1, 3 and 5". Plans land in `plans/` with an index and order; the source tree is left untouched.
A solo maintainer of an open-source library
Turn an audit into tasks contributors can pick up
Plans are self-contained Markdown, so they can be handed to a contributor or an agent as written. Publish them as issues with `--issues` only after reading each one, since security plans on a public repository become public.
A developer on a fixed model budget
Use the expensive model for review and a cheaper one for the edits
`/improve execute 001` dispatches a Sonnet executor in an isolated worktree (or `execute 001 haiku` for a smaller one), then re-runs the plan's checks and reports approve, revise or block. Merging stays manual.
Before you commit
- Setup
- `npx skills add shadcn/improve`, then `/improve` (or `/improve quick` for a cheaper first pass) inside the repository. The repository also ships a Claude Code plugin manifest.
- Integrations
- Any agent that supports the Agent Skills format; subagent fan-out and worktree execution are written for Claude Code. Uses `gh` for `--issues`. Reads ADRs, PRDs, CONTEXT.md and DESIGN.md when present so decided trade-offs are not reported.
- Data and privacy
- Read-only on your source: the only writes go to `plans/`, and secret values found in the audit are referenced by location and type, never copied. Your agent still sends code to its model provider. MIT licence.
- Cost at scale
- No paid tier. Cost is model usage: up to 4 concurrent subagents by default and up to 8 on `deep`, plus executor runs. No token or price figures are published.
- Limits
- Declines to implement anything itself. Fresh worktrees lack `node_modules`, so executors must install dependencies first. No releases; 23 commits, the latest on 12 September 2026.
What’s good
- Every finding cites a file and line, and is re-checked before you see it
- Plans carry pasted code, verification commands and STOP conditions
- Never edits source; executors work in disposable worktrees
- Cost scales with quick, standard and deep levels
What’s not
- No published token or savings figures
- Deep runs fan out to up to 8 subagents on your strongest model
- --issues can publish findings on public repositories
- No tagged releases; last commit September 2026
Best for
Teams on Claude Code or another skills-capable agent who want a strong model's code review turned into small, checkable tasks they or a cheaper model can execute.
improve alternatives
improve: common questions
- Is improve worth it?
- improve turns an expensive model into a careful reviewer that writes work orders instead of code, which is the right split, though nobody has published what the split saves. We rate it 3.9 out of 5.
- How much does improve cost?
- Free and open source (MIT). Model usage is billed by your coding agent's provider.
- What are the downsides of improve?
- No published token or savings figures; deep runs fan out to up to 8 subagents on your strongest model; --issues can publish findings on public repositories; no tagged releases; last commit September 2026.
- Who is improve best for?
- Teams on Claude Code or another skills-capable agent who want a strong model's code review turned into small, checkable tasks they or a cheaper model can execute.
- How much does improve cost for a team or in production?
- No paid tier. Cost is model usage: up to 4 concurrent subagents by default and up to 8 on `deep`, plus executor runs. No token or price figures are published.
- Is improve safe for company data?
- Read-only on your source: the only writes go to `plans/`, and secret values found in the audit are referenced by location and type, never copied. Your agent still sends code to its model provider. MIT licence.
- What are the best improve alternatives?
- The closest alternatives we have reviewed are Superpowers, BMAD Method and Open Code Review.
Answers checked Nov 20, 2026.