Launching today
ars
Learns your coding style and enforces it in new projects
1 follower
Learns your coding style and enforces it in new projects
1 follower
An engine that learns your coding style from past projects, turns it into agent skills, and proves coding agents follow it. Claude Code today; Codex and other agents next.
Coding agents write competent code that doesn’t look like the rest of your codebase. The usual fix is a hand-written conventions file, but you’re guessing at your own conventions, and nothing tells you whether the agent follows them.
ARS, an open source engine treats this as a measurement problem. It mines a reference repo for conventions: AST statistics for syntax and naming, the import graph for architecture, and an optional LLM pass for higher-level patterns. Each convention becomes an executable rule with evidence and exceptions (e.g. “hooks live in src/hooks/, 47/49 files”), and each rule tests itself by planting a violation and checking it’s caught. You review the rules; ARS generates Claude Code skills and an edit hook.
The same rules then score agent output on unseen tasks:
rebuilding held-out files from the repo
new apps in unrelated domains, with hidden acceptance tests
a blind judge picking which file is real
Each runs three ways: no skills, a generic guide of the same length (so gains aren’t just from a longer prompt), and your skills, with repeats and bootstrap CIs. Failing rules are diagnosed with counterfactual reruns, and fixes only land if a regression gate passes.
[Your real result here]
Limitations: TypeScript/React only for now, and the LLM pass can suggest plausible-but-wrong rules, which is why review is required.
Local-first, Apache-2.0. Codex and other agents are next. I’d especially love feedback on the measurement design.