Comments
>log in to comment👋 Hi everyone! I built ARS because AI coding agents write correct code, but not my code. Every new project meant correcting the same habits again, so I tried writing Claude Code skills for my style, and getting them right turned out to be really hard: You don't know your own style well enough to write it down. Most of it is habit: how you name hooks, where files go, how you handle errors. You only notice it when the agent gets it wrong. You can't tell whether a skill works. The agent reads it, then still does its own thing, and there's no way to measure whether the output improved or you just got lucky. Rules fight with "best practices." Generic advice wins over your personal choices unless the skill is precise and backed by examples. Skills get long and stale. Every fix adds another rule, the context fills up, and the code you wrote six months ago no longer matches what the skill says. ARS does that work for you. It learns your style from one of your projects, backs every rule with evidence from your code, and packages it as Claude Code skills. Then it tests whether the agent actually reproduces your style on tasks it has never seen, compared against runs with no skills and with a generic style guide. Install the skills into a new project, and the agent builds it your way from the first file. Where it is today: ARS supports React projects (TypeScript or JavaScript) and Claude Code. That's the starting point, not the limit. What's next: 🌍 More languages and frameworks: Python, Go and others. 🤖 More coding agents: Codex and others, compared side by side in the same lab. 🧠 Context window management: load only the rules and examples a task needs, so the skills stay small as your style grows. 🌳 Git worktrees for agents: each agent works in its own worktree, so several can run side by side on a real project. 🤝 Collaborators wanted. ARS is open source (Apache-2.0) and early. If you'd like to add a language or an agent, improve the analysis, or just try it on your code and tell me where it gets your style wrong, you're very welcome. Open an issue, send a PR or reply here. I'd love your thoughts on the workflow, and on what you'd want ARS to support next. https://github.com/ars-engine/ars
ARS learns your React code style, creates Claude Code skills, and verifies agents follow your conventions.
- for
- Developers who want AI agents to code in their personal style for React projects.
- pricing
- open source
- license
- Apache-2.0
- v0.1.1Security release. If you use 0.1.0, upgrade.3 days ago
- v0.1.0v0.1.0 — first public release5 days ago
Key features
- Signature extraction — Analyzes a React codebase to learn formatting, idioms, naming, component anatomy, architecture and voice.
- Review inbox — Allows you to approve, reject or reword each learned rule.
- Versioned skill package — Generates installable SKILL.md files, formatter config and edit hooks for projects.
- Lab testing — Runs agents with and without skills on unseen tasks and reports confidence intervals.
- Improvement loop — Diagnoses failing rules, proposes fixes and promotes only validated improvements.
- Local analysis UI — Provides a web UI to paste a repo URL or folder, analyze locally and view results instantly.
Use cases
- Generate a skill package so Claude Code writes new React components in your team's idioms
- Validate that an AI coding agent adheres to your project's naming and architecture conventions
- Iteratively improve style rules by reviewing failed cases and applying fixes
- Create a reproducible coding style guide that can be shared across multiple React projects
Ars vs alternatives
Cursor | Tabnine | OpenCode | |||
|---|---|---|---|---|---|
| Best for | Style-aware coding agents for React | AI-assisted editing | Team code completion | Context-rich coding agents | Open-source terminal agents |
| Pricing | Open source | Subscription | Subscription | Subscription | Free |
| DevHunt upvotes | 1 | 0 | 0 | 0 | 0 |
| Launched | Nov 2026 | — | — | — | — |
- Ars vs Cursor: Cursor is an AI-powered code editor, while ARS focuses on learning and enforcing a specific coding style for agents.
- Ars vs Tabnine: Tabnine provides AI code completions for teams, but does not generate verifiable style-specific skill packages.
- Ars vs Augment Code: Augment Code offers deep codebase context for agents, yet ARS adds a verification lab to prove style adherence.
- Ars vs OpenCode: OpenCode is an open-source terminal coding agent; ARS adds style learning and testing for React projects.
Ars FAQ
What languages does ARS support?+
Currently only React projects written in TypeScript or JavaScript are supported.
Do I need an API key?+
Claude is optional for analysis but required for running agents and some rules; you must provide an API key.
Is my code sent to external services?+
ARS keeps all data locally, only sending redacted excerpts to Claude and never stores tokens.
How do I install ARS?+
Clone the repo, run npm install, then npm run dev for the API and UI, or npm start to serve both together.
Can I use ARS with private repositories?+
Yes, ARS clones private repos using your machine’s git credentials and stores data only on your disk.
What is the improvement loop?+
When a rule fails, ARS diagnoses the cause, suggests a fix, and promotes the fix only after it passes regression tests on held-out tasks.
Summarized by DevHunt from github.com · Oct 30, 2026. Details may change; check the official site.
Cursor
Tabnine
OpenCode










