My bots started arguing with each other.

Claude wanted useEffect. Gemini kept steering toward a custom hook. I just needed a button to be orange. I spent twenty minutes in the middle of an argument between two language models about React philosophy while my son waited to do his maths homework.

The backend situation was worse. I'd been using different AI assistants across different sessions without giving any of them the actual database structure. So they each invented their own version. One thought users had an id. Another was writing queries against user_uuid. A third had added a profile_id field that didn't exist anywhere in the real database.

Nothing was malicious. Each AI was doing its best with incomplete information. But "incomplete information" is the default state in a multi-model workflow if you're not deliberate about it.

The fix was almost embarrassingly simple. I wrote up the schema — every table, every column — and pasted it at the top of every session. Not as background reading. As a contract. Here is what exists. Work within this.

The column name disagreements stopped. The frontend pattern wars calmed down once I did the same for component conventions.

Managing AI assistants is stakeholder management. Give everyone the same briefing document before the meeting starts, or they'll all show up working from a different version of the plan.