improve: shadcn's skill where a strong model plans
Added NewCopyable prompt below
Published Last verified Released
Our verdict
A practical way to spend your expensive model where it pays off (understanding the repo and deciding what is worth doing) and let cheaper models do the typing. /improve maps the repo, fans out subagents across nine audit categories, re-reads every cited file:line itself to drop false positives, and ranks findings by impact over effort. You choose which become plans in plans/, each with code excerpts, the repo's own test and lint commands as checks, and stop conditions for a smaller model. It has been public since June 2026 and is back on GitHub Trending this week (about 11,200 stars). Builders report real use: Luke Berry's step-by-step post (365 likes) and @awpthorp "ploughing through issues" with it. Limits: a full or deep audit with parallel subagents burns a lot of tokens on your best model (use /improve quick first); execute needs an agent that can spawn subagents and git worktrees; plan quality depends on the model you run it with; and nobody has published measured savings. We have not run it ourselves.
- Who it's for
- Developers and indie makers using Claude Code, Codex, Cursor or another Agent Skills-compatible agent on a real codebase, especially those paying for a top model and wanting to stretch it
- Leverage
- A prioritised, evidence-backed backlog and ready-to-run specs from one command, so cheaper models or teammates can execute fixes without re-reading the codebase
- Effort
- Low
- Date added
What it does
improve (shadcn/improve, MIT) is an Agent Skills-format skill, also packaged as a Claude Code plugin. Recon maps the stack, conventions and exact build/test/lint commands, and reads ADRs, PRDs and design docs so settled trade-offs are not re-flagged. Audit covers correctness, security, performance, test coverage, tech debt, dependencies, developer experience, docs and direction (feature ideas that must cite evidence from the repo). Findings carry file:line evidence, impact, effort and confidence; rejected ones are recorded so they do not return. Plans are written for the weakest plausible executor: all context inlined, a verification command per step, out-of-scope lists, and the git commit they were written against. execute runs a cheaper executor in an isolated worktree and reviews its diff (approve, send back up to twice, or block); reconcile refreshes the backlog; --issues publishes plans as GitHub issues. It only writes to plans/ and never reproduces secret values.
How builders use it
- Install with npx skills add shadcn/improve, open your agent in the repo and start with /improve quick to keep the first run cheap.
- Reply with the findings you want planned ("plan 1, 3 and 5") and review the files written to plans/.
- Hand a plan to a cheaper agent ("implement plans/001-*.md") or run /improve execute 001 and merge only after reading the verdict and diff.
- Run /improve branch before opening a PR, and /improve reconcile at the start of the next session.
What to expect
Your top model's time goes into judgement and specs instead of edits, and each fix arrives as a reviewable plan with built-in checks, so smaller models can finish the work with fewer mistakes.
Copy this: Install + first audit
# Official source: github.com/shadcn/improve (MIT)
npx skills add shadcn/improve
# In your agent, inside the repo:
/improve quick # cheap pass: hotspots, top findings only
/improve # full audit -> findings table -> plans
/improve security # focused audit (also perf, tests, bugs)
/improve branch # only what this branch changes
/improve execute 001 # cheaper executor in a worktree, then review
/improve reconcile # verify, unblock, retire old plans
# Hand a plan to any other agent:
Implement plans/001-<name>.md exactly. Run every verification command it lists and stop if a STOP condition is hit.Related discoveries
See the newest picks →- ToolCoding & Shipping Agents
Plannotator: review agent plans and diffs in your browser
An open-source (Apache-2.0), local, browser-based review surface for coding agents. When Claude Code, Codex, Copilot CLI, OpenCode, Pi and others propose a plan or finish a change, Plannotator opens it in your browser so you can mark it up and send structured feedback straight back to the agent.
NewEffort: LowPrompt#Plugin / extension#Open source
- WorkflowCoding & Shipping AgentsHot
Run a codebase-wide audit with Claude Code dynamic workflows
Have Claude write a script that fans an audit out to many subagents, cross-checks findings, and can be saved and rerun.
Effort: MediumPrompt#Emerging
- ToolCoding & Shipping Agents
Cloudflare security-audit skill: agent bug hunts
Cloudflare's MIT skill that turns Claude Code, Codex or Cursor into a six-phase security auditor where a fresh agent tries to disprove every finding.
Effort: MediumPrompt#Plugin / extension#Open source
- ToolCoding & Shipping Agents
Elephant-Goldfish Model: specs reviewed by fresh agents
Google's Apache-2.0 plugin that makes your coding agent draft a design spec with you, then has fresh zero-context subagents attack it before any code is written.
Effort: MediumPrompt#Plugin / extension#Open source
Get the best new A.I. picks in your inbox
The best new agent workflows and tools we’ve vetted, with a clear verdict on each. Short, occasional, useful.