A new open-source agent skill from shadcn, the developer behind the widely used shadcn/ui component collection, inverts the usual way people spend money on AI coding. Instead of pointing a cheap model at a task, /improve uses your most capable - and most expensive - model for the part where intelligence compounds: understanding the codebase, judging what is worth doing, and writing the specification. Execution goes to cheaper models.
The skill, installed with a single npx skills add shadcn/improve command and released under the MIT licence, never implements anything itself. Run /improve in a repository and it maps the repo's stack, conventions and exact build/test/lint commands, then fans out parallel subagents across nine categories - correctness, security, performance, test coverage, tech debt, dependencies and migrations, developer experience, docs, and direction, meaning feature suggestions that must cite evidence from the repo. Every finding carries a file:line reference, impact, effort and confidence. It then re-reads each cited location itself to drop false positives before showing anything, and writes one self-contained plan per selected finding into a plans/ folder.
The plans are written for the weakest plausible executor: a model that never saw the advisor session and may be much smaller. Context is inlined, every step ends with a command and its expected output, and each plan carries explicit out-of-scope lists and STOP conditions rather than leaving a small model to improvise. Plans stamp the git commit they were written against so an executor can run a mechanical drift check first.
There is a follow-through loop too. /improve execute <plan> spawns a cheaper executor in an isolated git worktree, hands it the plan, then reviews the result like a tech lead - re-running every done criterion and checking scope compliance - and returns a verdict of approve, revise (up to two rounds) or block. /improve reconcile then verifies that finished plans still hold, rewrites blocked ones around the obstacle, refreshes drifted plans and retires findings that were fixed independently. Variants scope the audit to a single branch, to security or performance only, or publish plans straight to GitHub issues.
The repository's hard rules are as notable as its features: it never modifies source code, it never runs commands that mutate the working tree, it never reproduces secret values (locations and credential types only, with rotation always recommended), and if asked to implement something it declines and points at the plan. The repo trended among GitHub's fastest-growing projects this week. As agents write more code, the question of which model should be doing which part is becoming an engineering decision in its own right.

