The problem
Every AI coding tool has a context problem. Claude Code reads CLAUDE.md. Codex reads AGENTS.md. Gemini reads GEMINI.md. But none of them share context, and none of them scale past a single file in a single repo.
The result: your engineer's Claude knows the codebase but not the product. Your CEO's Claude knows the product but not the stack. New hires get no context at all. Everyone re-explains the same things, every session, every day.
What we tried first
ETH Zurich researchers found that auto-generated context files reduce task success by 2-3% and increase cost by 20%. Generic context makes AI worse, not better. We tried everything else too:
- Notion/Confluence wikis - AI can't read them. Nobody updates them.
- Shared Google Docs - not in the AI's context window. Goes stale within a month.
- Slack channels for "AI tips" - drift immediately. Not version-controlled.
- Symlinks between config files - hacky, breaks on some systems, doesn't scale.
- One giant CLAUDE.md - 2,000 lines. Nobody maintains it. Goes stale in 2 weeks.
Without antidrift
- Re-explain your company every session
- Copy-paste from Google Docs into Claude
- One person's Claude is smart, everyone else's is dumb
- New hires spend a week asking "how does this work?"
- Tribal knowledge walks out the door when someone leaves
With antidrift
- Every session starts with full context automatically
- Context lives in git, not docs nobody reads
- Every teammate's AI knows the same things
- New hires run
/onboardand get walked through everything - Knowledge survives in the brain, not in heads
The only thing that worked: plain markdown files in a git repo, organized by department, maintained by the people who own that knowledge.
How it works
When you run antidrift init, it creates a brain - a git repo with directories for each part of your company. Each directory has a CLAUDE.md, AGENTS.md, and GEMINI.md that the corresponding AI tool reads automatically.
product/
CLAUDE.md · AGENTS.md · GEMINI.md
engineering/
CLAUDE.md · AGENTS.md · GEMINI.md
marketing/
CLAUDE.md · AGENTS.md · GEMINI.md
sales/
CLAUDE.md · AGENTS.md · GEMINI.mdHere's what a real context file looks like:
# Engineering
## Stack
Next.js 14, PostgreSQL, Redis, deployed on Vercel.
## Conventions
- Functions under 30 lines
- Tests required for all API routes
- Feature branches, squash merge to main
## Current sprint
Migrating auth from NextAuth to Clerk.
Don't touch /lib/auth until March 15.This is what your AI reads on every session. Not a 2,000-line dump - structured, scoped, maintained by the people who own it.
Why directories: Different people own different context. Engineering owns the stack. Marketing owns the messaging. Nobody merges conflicts on a single file. Each directory is a natural git boundary.
Why three files: Claude Code reads CLAUDE.md. Codex reads AGENTS.md. Gemini reads GEMINI.md. Antidrift's cross-compiler keeps all three in sync - edit one, run /push, and the other two update automatically.
How it stays current
The brain updates when people use it - not when someone remembers to update docs.
- /push - commits and syncs changes. Cross-compiles all context files before pushing.
- /ingest - imports existing docs, wikis, or files into the brain. Already have a Confluence wiki or a Google Doc handbook?
/ingestbrings it in. - /refresh - pulls the latest brain from the team. Stashes local changes, rebases, and restores.
- /onboard - walks a new hire through the entire brain interactively. Day one context, not day thirty.
Already have a CLAUDE.md? Antidrift wraps around it. Already have 40 repos? The brain is a separate repo - it doesn't live inside your code. Works for teams of 2 or 60.
The brain lives in git, so you get version history, diffs, and blame for free. When someone changes the product roadmap, you can see who changed it, when, and why.
Try it
npx @antidrift/cli init One command creates a brain. Your next AI session starts with full company context.