
Improve
github.com/shadcn/improve- Category
- Developer Tools
- Rank
- No. 77Tools index
Previous survey · No. 85 ·
- Type
- APP
- Builder
- @shadcn
- GitHub
- 9.1k stars
- Date
About
Use your most capable model to audit your codebase and write plans for cheaper models to execute.
What it does
Improve surveys a repository, checks findings against cited code, and turns selected work into standalone Markdown specifications. Each specification includes context, scope boundaries, ordered steps, verification commands, expected results, drift checks, and stop conditions for a separate executor.
Why it's ranked here
Its strongest idea is disciplined separation between judgment and implementation. Findings require file-and-line evidence, the advisor rechecks reported locations, and users choose which items become plans. The follow-through is unusually complete: it can track drift, refresh stale work, dispatch an isolated executor, and verify the resulting diff without merging it.
What's good
Plans target agents with no prior session context. They inline current code, repository conventions, test commands, dependencies, exclusions, and machine-checkable completion criteria. The audit covers correctness, security, performance, tests, architecture, dependencies, developer experience, documentation, and product direction. Rejected findings are recorded, reducing repeated false positives during later audits.
Tradeoffs
The advisor deliberately refuses direct source edits, so it cannot replace an implementation agent or finish a change by itself. Its value depends on the host supporting Agent Skills, and automated execution additionally depends on subagents, Git, and isolated worktrees. Thorough audits can fan out across several specialist agents, which increases model usage before implementation begins.
How to use it well
It suits teams that want careful technical triage followed by explicit handoffs to agents or humans. Start with a repository audit, select only worthwhile findings, inspect the generated specifications, then assign them to executors. Use branch-focused analysis before pull requests and backlog reconciliation between sessions. Choose another tool when you need direct editing, merging, or deployment.
Technical notes+
The main behavior is defined in skills/improve/SKILL.md, with planning structure in skills/improve/references/plan-template.md, audit criteria in skills/improve/references/audit-playbook.md, and executor, reconciliation, and issue flows in skills/improve/references/closing-the-loop.md. README.md documents installation through npx skills add shadcn/improve and slash-command variants. .claude-plugin/plugin.json and .claude-plugin/marketplace.json provide Claude plugin metadata. examples/001-extract-shadow-config-resolution.md demonstrates the generated plan format, including a commit-based drift check, scoped files, verification gates, done criteria, and STOP conditions.
Observed
- License
- MIT
- Primary format
- Markdown agent skill and Markdown implementation plans, with JSON plugin metadata
- Install surface
- Installed with npx skills add shadcn/improve
- Interface
- Agent Skill with slash-command variants for audits, planning, execution review, reconciliation, and GitHub issue publication
- Platform support
- Works in agents that support the Agent Skills format
- Source behavior
- The advisor writes only planning artifacts; separate executors may edit code in isolated Git worktrees
- Repository structure
- Includes a sample generated plan plus separate audit, plan-template, and closing-the-loop references
Read from README.md, LICENSE.md, .claude-plugin/plugin.json, .claude-plugin/marketplace.json, examples/001-extract-shadow-config-resolution.md, skills/improve/SKILL.md, skills/improve/references/plan-template.md, skills/improve/references/audit-playbook.md, skills/improve/references/closing-the-loop.md.
Intel on Improve
- A Removal Based Approach to Improve LLM Faithfulness at Test-Time
- Run this to install: npx skills add --skill improve-threejs Powered by React Doc
- Introducing /improve-threejs Turn vibe-slop Three.js prototypes into stunning…
- Verified Tool Calls Improve LLM Agent Reliability Under Non-Atomic Failures
- (Im)Paired Programming: Coding Agents Improve Productivity but Harm Understanding
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.