Use improve to audit a codebase and emit only plans

Who it is for

If you want a stronger model to see the codebase clearly, decide what is worth doing, write a spec, then hand execution to a cheaper model, improve fits. Officially it is written as: audit any codebase and write implementation plans for other agents. The skill itself implements nothing; the plan is the deliverable.

The first run should stop at the plan. It will not change source. The only allowed write is plans/. If you want the same skill to change product code directly, you picked the wrong one. The executor should only change a disposable worktree; you always do the merge. This page gives no worktree commands; do not invent them.

Public docs say it works in any agent that supports the Agent Skills format. Plans are ordinary markdown that an agent or a person can continue. No specific client is named. Do not add a product roster, and do not call it exclusive support.

First try in 3 minutes

Use only this official install command:

npx skills add shadcn/improve

Prerequisite: you have a real codebase to audit and can run npx in that project. After install, be ready to use /improve. Do not invent a second install method.

Install writes skill files. Before running it, confirm the current directory is the repository you want to audit. After the command, first do a check that writes no plan: confirm the agent recognizes /improve and can restate "only write plans/, do not change source". That is the observable success signal for install.

A real /improve run does a full audit, ranks priority, and writes plans into plans/. Confirm you accept new markdown plan files, then run:

/improve

Success signal: source is untouched by it, and plans/ contains prioritized findings and implementation plans. Public material lists no other slash commands and does not guarantee plan-file naming.

Use cases

Case 1: Full audit of the current repo, then a plan. With it installed, run /improve and let it finish audit → priority findings → plan. Read the markdown in plans/ first, then decide whether to hand it to someone else. This is the skill's main workflow. Public docs do not give an audit-dimension checklist; do not pretend you already know which files it will inspect.

Case 2: Hand the plan to another agent or person. Officially the plan is ordinary markdown anyone can pick up; this skill still implements nothing. The executor should only change a disposable worktree; you finish the merge. Confirm the plan scope first, then hand it over. This page gives no execute or merge commands. This case only covers "hand off the plan and keep merge control yourself".

Limits

The skill itself never changes source. Writes go only to plans/. Do not treat it as auto-refactor or auto-fix. Do not let it execute the plan in the main workspace.

Install is only npx skills add shadcn/improve. The page names no client, only support for the Agent Skills format. Do not add marketplace, and do not copy a skills directory that was not given.

/improve writes plan files; confirm that before you run it. Public material has no uninstall steps and does not turn "stronger model / cheaper model" into required concrete model IDs.

Repository URL:

https://github.com/shadcn/improve