improve
Use your most capable model to audit your codebase and write plans for cheaper models to execute.
What is it?
What it is
An agent skill that audits any codebase and writes self‑contained implementation plans for other agents (or humans) to execute.
Why it exists
To use a powerful model for high‑value reasoning—understanding the codebase, judging what’s worth doing, and writing specs—while delegating actual code changes to cheaper, faster models, keeping the skill itself from writing code.
Who should use it
Developers and tech leads who use agent-based workflows and want automated codebase audits and actionable plans for implementation by cheaper models or humans.
Who should avoid it
Teams that do not use agent skills, prefer fully manual code review, or have no need for automated planning; also beginners unfamiliar with agent frameworks.
How it works
A quick walkthrough in plain English
How improve works
Step 1 of 3
You interact with it
Open improve, send a request, or connect it to your stack.
Features
Advantages
- Uses expensive model only for high-value planning/intelligence
- Delegates execution to cheaper, faster models
- Never modifies source code directly; only writes to plans/
- Plans are self-contained and usable by any agent or human
- Verification gates make success machine-checkable
- Reduces false positives via advisor re‑read of cited locations
- Leverages existing repo documentation to avoid re‑flagging decided tradeoffs
- Prioritized backlog focuses effort on highest leverage items
- Works with any agent that supports Agent Skills format
- Outputs plain markdown, easy to review and version
- Provides explicit out‑of‑scope and STOP conditions to prevent model improvisation
- Supports iterative reconciliation to keep plans up‑to‑date
- Can publish plans as GitHub issues for existing workflows
Disadvantages
- Requires an agent environment that supports the Agent Skills framework
- Initial advisor run may need a costly large‑language model
- Plannered plans may not for highly nuance
- Relies on the executor model to correctly follow verification gates; failures may require manual intervention
- Plan files can become verbose due to inlined context and examples
- No direct code changes; merging and shipping still require human or separate automation
- Limited to the predefined audit categories; custom categories would need extension
- Dependent on the presence of repo conventions and test/lint commands for verification gates
Installation
native
npx skills add shadcn/improve
FAQ
What is improve?
improve is an open-source project licensed under MIT.
What language is improve built with?
Primary language: multi-language.
Is it free to use?
Yes. Licensed under MIT. Check the license for commercial use.
Featured in Videos
YouTube tutorials and walkthroughs for improve
Alternatives
Similar projects ranked by category, topics, and text overlap.