LabscoMCP early access ↗
REPO PACKAGECOMMUNITY

shadcn improve

shadcn · publisher6,953 GitHub starsMIT · Listed as freegithub.com/shadcn/improve

LABSCO SUMMARY

Summary

A single skill, and an unusually well-specified one: it audits a codebase, writes implementation plans for another agent to execute, and refuses to touch your source code itself.

The skill (shadcn/improve, installed with npx skills add shadcn/improve) runs a four-stage pipeline behind one command, /improve. Recon maps the repo's stack and its own build, test, and lint commands; a fan-out of subagents then audits nine categories — correctness, security, performance, test coverage, tech debt, dependencies, developer experience, docs, and feature direction — and every finding must cite a file and line from the actual code rather than a generic suggestion. A second pass re-reads each cited location before anything is shown, specifically to drop false positives. What survives lands in a table ordered by impact over effort, and only the findings you pick become plan files under plans/.

This fits teams already comfortable splitting an expensive "thinking" model from a cheaper "doing" one. The plans it writes are explicitly built for a smaller executor that never saw the audit session — inlined context, machine-checkable verification steps, and stated STOP conditions in place of judgment calls. If your workflow already uses one model for both auditing and fixing, the split this skill is built around adds a step you may not want.

READ MORE

Read-only is a stated hard rule, not just a claim.

The README lists it directly: the skill itself never edits source and only writes to plans/. Its execute subcommand hands a plan to a separate executor in an isolated git worktree — that executor is the one that changes code, with the audit skill reviewing the diff afterward rather than writing it.

The one worked example is the author's own.

The README's sample findings, and the linked example plan, come from a run against shadcn/ui — a repository the same author maintains. It is a real, checkable example, but it is not an independent test of the skill.

Unofficial, one skill, no outside track record we can see.

There is no companion collection, no listed users, and nothing in the repository confirming how the review-and-revise loop (execute, reconcile) behaves once a plan is handed off in practice — we have read what it claims to do, not watched it run.