shadcn/improve

What it solves

It addresses the inefficiency of using high-cost, high-intelligence AI models for repetitive coding tasks. Instead of having one expensive model do everything, this tool separates the high-level "intelligence" (auditing and planning) from the low-level "execution" (writing the code), allowing cheaper models to handle the implementation phase.

How it works

The tool operates as an agent skill that follows a specific pipeline:

1. Recon: It maps the repository's stack, conventions, and documentation (like ADRs or PRDs) to understand the project's context.
2. Audit: Parallel subagents scan the codebase across nine categories (including security, performance, and tech debt) to find issues, citing specific file and line evidence.
3. Vet & Prioritize: An advisor model filters out false positives and presents a prioritized table of findings based on impact and effort.
4. Plan: For selected findings, it generates self-contained Markdown specifications in a plans/ directory. These plans include inlined code excerpts, machine-checkable verification gates, and explicit stop conditions so that even small, "weak" models can execute them without hallucinating.
5. Execute & Reconcile: It can dispatch cheaper executors to implement plans in isolated worktrees, review their diffs, and periodically refresh the backlog to account for code drift.

Who it’s for

It is designed for developers and teams using AI agents to maintain and improve their codebases who want to optimize for both cost and quality by using a "tech lead" model to guide "junior" executor models.

Highlights

• Separation of Concerns: The skill never modifies source code directly; it only writes plans.
• Machine-Checkable Plans: Every plan includes specific commands and expected outputs to verify success.
• Isolated Execution: Uses disposable git worktrees for implementation to keep the main branch safe.
• Context-Aware: Ingests design docs and ADRs to ensure suggestions align with existing product intent.