Collaboration is what happens after the first answer. Instead of treating the model as a vending machine, you treat it as a colleague who has read nothing about your situation: you supply the material, it drafts, you react to specifics, and the two of you converge — with the model doing the typing and you doing the judging.
How the loop works
- Give it the room to work in. Documents, data, code, prior decisions. A model with your context beats a bigger model without it.
- Draft, then react. Ask for a draft early and correct concretely — “rewrite section 2 as three benefit statements”, not “this is off”.
- Split the roles. Let the model generate options and do the legwork; keep taste, priorities and final judgement on your side.
- Keep the thread clean. Prune contradictions as you go; an argumentative history degrades every later answer.
Ways to divide the work
- Draft and critique. One pass writes, a second pass reviews against your criteria and lists only what fails.
- Grounded conversation. Retrieval puts your own documents into the exchange so answers cite your sources — see Embeddings & RAG.
- Mixed media. Screenshots, diagrams and recordings are legitimate input for a multimodal model — see Multimodal Systems.
- Shared standards. A reusable house-style block pasted into every session keeps output consistent across people.
Example
Vague
That draft isn't right, do it again.
Better
Keep: the structure and the opening paragraph. Fix: section 2 is feature-led; I need customer benefits. Add: the migration timing from the attached plan. Rewrite section 2 only, three sentences, and list any fact you had to assume.