// the find
shadcn/improve
Use your most capable model to audit your codebase and write plans for cheaper models to execute.
An Agent Skill (not a library, just prompt files + workflow logic) that runs a full-repo audit with a capable model, then writes self-contained markdown plans in plans/ for a cheaper model to execute in an isolated worktree. Aimed at teams already using agentic coding tools who want to stop burning an expensive model's tokens on rote implementation.
Verification gates baked into every plan step (repo's own build/test/lint commands plus expected output) directly targets the real failure mode of small-model executors: not being able to judge their own success. The vet stage explicitly re-reads every subagent-cited file:line before surfacing findings, which is a real answer to the known problem of parallel audit subagents over-reporting. Hard boundaries are enforced structurally, not just claimed — audit is read-only, only plans/ gets written, executors only touch disposable worktrees. Ingesting ADRs/PRDs/CONTEXT.md so it doesn't re-litigate already-decided tradeoffs is the kind of detail that comes from actually running this against real repos, not a demo.
There's no code here — it's SKILL.md plus reference markdown, so the 'audit' and 'vet' stages are entirely prompt behavior with no eval suite proving they hold up across repos or model swaps; quality is whatever the underlying model does that day. The leverage-ranking (impact/effort weighted by confidence) is asserted but never shown failing or being second-guessed, which matters for a tool whose whole pitch is 'don't trust vibes.' It's locked to the Agent Skills format via npx skills add, so portability depends entirely on how many agents actually adopt that spec. Running parallel subagents across nine audit categories on `/improve deep` for a large repo has an obvious cost profile that's never addressed — no token/cost estimate or budget guard mentioned anywhere.