Key takeaways
- Let your most capable model audit your codebase and write the plans, so cheaper models handle the typing.
- Findings are re-checked: the advisor re-reads every cited location before showing you anything.
- Plans are built for the weakest executor, with verification steps and clear stop conditions.
- You stay in charge: the skill never implements anything itself, and merging stays up to you.
- The catches: subagents over-report, the README publishes no measured savings, and you need a compatible agent.
What it is
improve describes itself as an agent skill that audits any codebase and writes implementation plans for other agents to execute. It is open source under the MIT licence, and the code is on GitHub at shadcn/improve.
The problem it solves
Developers who use AI agents face a trade-off. The strongest models are best at judging what is worth doing, but running them on every change is costly. The project's idea is to use your most capable model for the part where intelligence compounds, and hand execution to cheaper models.
The skill never implements anything itself. The plan is the product.
What you get
- Broad coverage: the audit fans out parallel subagents across nine categories.
- Findings you can check: each one carries evidence, impact, effort and confidence.
- Fewer false alarms: the advisor re-reads every cited location before showing you anything.
- Plans a small model can follow: they are written for the weakest plausible executor, and every step ends with a command and its expected output.
- Optional hand-off: the execute command dispatches a cheaper model in an isolated worktree.
The catches
- Over-reporting: the README itself says subagents over-report, which is why the vetting step exists. Expect to review the findings table yourself.
- No published savings: the README makes the case for cheaper execution but gives no measurements of cost or quality.
- Compatibility: it works in agents that support the Agent Skills format, so it is not a standalone tool.
- Review still matters: plans include explicit out-of-scope lists and stop conditions, but merging stays up to you.
What we did not test
We did not install the skill, run an audit or execute a plan, so we cannot say how accurate the findings are or how much money the approach saves. Every claim here comes from a literal quote in the project's README or the GitHub API, and the feature descriptions are the project's own.
Verdict
Editors' view: this suits developers who already work with coding agents and want a written, reviewable plan before any code changes. If you prefer to let one agent audit and edit in a single pass, it adds a step. If you try it, start on a small repository and read the plans critically.
Sources
Need help adopting this, or have a software project in mind?
Tell us what you are building. The initial consultation is free. Email [email protected] or follow @zolnaltd.
Start a conversation →