coldteadotai/abide

Play the video in the post below

Abide checks a coding agent against project rules that a linter can’t express, things like “never let a raw error reach a user” or “don’t create premature abstractions.” Every AGENTS.md or CLAUDE.md is full of instructions like these that nothing enforces. Ohans Emmanuel built Abide to enforce them anyway, on every edit or turn the agent makes.

For each rule, Abide asks Jev one typed question about the diff and gets a probability back. Jev sees only the rule and the diff, never the running conversation, so the check is the same on edit 200 as on edit 1. When a rule is broken, the agent is told which one and fixes it in the same turn, before moving on. Rules a linter can already check are left to the linter; Abide only takes the ones that need judgment.

Emmanuel replayed 93 real Claude Code sessions, 1,256 edits and 147 turns across two repositories, against their own AGENTS.md files, for a total of 22 cents. The agent broke a rule on 1 turn in 13, starting from its first edit. Jev flagged 39 edits and 15 turns as violations; an independent human reviewer confirmed 10 of the edit-level flags and 11 of the turn-level ones, with the turn-level catches, things like single-use abstractions and oversized files, holding up 11 times out of 15. Each check runs in about 300ms for a few thousandths of a cent, which Emmanuel argues makes this kind of per-edit review practical for the first time: the same check with an ordinary language model would cost a cent or more and take several seconds per edit, then need its prose answer parsed. Abide is open source and works with Claude Code, Codex, and opencode.

Open on GitHub