TrendWhat rose on GitHub, every day at 07:30 KST · 한국어

GitHub · as of November 15, 2026 · View on GitHub →

shadcn/improve

AI coding agent skill

trending one of the last 30 days

Separate planning from execution to reduce costs while maintaining quality

The most capable model analyzes the codebase, prioritizes issues, and documents the plan, while a cheaper model or a human carries out the actual work based on that document. This division of labor assigns complex judgment to expensive models and repetitive execution to affordable ones, reducing costs without sacrificing quality. The tool does not modify code directly; instead, it produces a detailed blueprint specifying what to change and how.

The plan is the deliverable

The tool does not modify code or add features. Instead, it analyzes the repository, organizes the issues it finds, and generates a specific execution plan for each issue as a Markdown file in the plans/ folder. These plans are written to be independently readable, including the current state of the code, the file paths to be changed, and the commands for verification. As a result, another agent or developer receiving this document can start working immediately without reloading the context. The user always makes the final decision on whether to merge the code.

From analysis to execution

Users can request a full audit or a focused audit on specific areas (such as security or performance) by entering the /improve command. After understanding the repository's structure and conventions, the tool runs sub-agents divided into nine categories in parallel to find issues. It does not simply list problems; it assigns each finding a file location, impact, effort level, and confidence score. It then writes a detailed plan for the items selected by the user. Through the /improve execute command, it can also run a cheaper model in an isolated worktree to review the results.

Conditions for an executable plan

The generated plans follow three principles so that even the most limited executor can understand them. First, all context is included directly in the document, excluding vague expressions like "as discussed above." Second, the completion criteria for each step are specified with mechanically verifiable commands and expected outputs. Third, a clear boundary is set to instruct the executor to "stop and report" if the plan does not match the actual code, preventing arbitrary judgment. Additionally, the git commit at the time the plan was written is recorded to automatically verify whether the code has changed before execution.

Pre-use considerations

Released under the MIT license, the tool focuses solely on read-only analysis and plan creation without modifying source code. However, the plans generated by the tool must be reviewed by a human, and the authority to merge execution results rests with the user. If the repository contains ADRs or design documents, the tool references them to avoid re-flagging already decided matters as issues, making it more effective in projects where such documents are well-maintained.

Written by AI from this repository's README on November 11, 2026. For the facts, the original on GitHub is the reference.

By the numbers

GitHub trending history
1 day between Jun 19, when Trend began recording, and Nov 15, 2026 · first on Nov 11 · best daily rank #4 (Nov 11)

Read from GitHub, as of November 15, 2026.

Stars
11,944
Contributors
5
Forks
481
Commits
23
Last commit
September 12, 2026
Open issues
6
Open pull requests
12

View on GitHub →

README badge

trend.ondo.sh · GitHub daily trending #4 · November 11, 2026

The badge above shows the highest this repository has ranked on GitHub's daily trending since June 19, 2026, and the day it did. Paste the Markdown below into your README. The badge opens this page, and it updates by itself when the repository climbs higher.

[![trend.ondo.sh](https://trend.ondo.sh/github/r/shadcn--improve/badge.en.svg)](https://trend.ondo.sh/github/r/shadcn--improve/en.html)