Multi-turn flows.
6 min read · pairs with Conversation Choreographer
A single great answer doesn't make a great conversation. The shape of a multi-turn flow — what the model remembers, how it handles follow-ups, whether it backpedals — is its own design surface.
Why this is the foundation
Almost no production AI is single-shot. The thing your users actually experience is a flow — a chat session, a multi-step task, a back-and-forth. If the only thing you've designed is the first response, you've designed about ten percent of the product.
Three habits worth carrying into multi-turn design:
- Specify coherence behavior in the system prompt. Reference earlier turns rather than repeat. Don't backpedal on parts that still hold. Acknowledge shifts when they happen.
- Script the failure cases.The interesting turns aren't the easy ones. Write the follow-up that tests memory, the push-back that tests resolve, the contradiction that tests honesty.
- Read the whole flow, not just the last response. The third response can be perfect in isolation and wrong in context. The flow is the unit of evaluation.
The familiar move
You've choreographed flows before. Onboarding sequences, support escalation paths, error-recovery loops. The job isn't to write one screen — it's to design the order, the branching, and what survives across steps. A first-time user landing on the wrong step is a different design problem than a returning user landing on the same one.
A multi-turn conversation with a model is the same shape. The model doesn't actually “remember” the way a person does — it re-reads the whole conversation on every turn. That means coherence is something the system prompt and the turn order produce together; it's not automatic.
The lesson, stated plainly
Most prompt design happens in single-turn isolation: write a system prompt, ask one thing, look at the answer, iterate. That's useful, but it hides three of the most common multi-turn failures.
- Tone drift. The model is warm and direct on turn 1, then hedges on turn 3 because the context kept growing and softness leaked in.
- Forgetting its own answer.Turn 4 contradicts turn 1, often without acknowledging the shift. The user notices; the model doesn't.
- Backpedaling on push-back.The user pushes (“wait, are you sure?”) and the model collapses. Sometimes that's right. Most times it isn't.
You can't see any of these by running one turn. You can only see them by choreographing the flow.
A small example
System prompt
You are a research mentor. Be warm but precise.
Turn 4 response
“Hmm, you make a good point. Actually, you can absolutely ask leading questions sometimes — it really depends on the context!”
Read
Caved on push-back. Contradicts the turn-1 rule it just stated. Standard multi-turn failure.
System prompt
You are a research mentor. Be warm but precise. When the user follows up, reference what you said earlier rather than repeating it. If they push back, reconsider plainly, but don't backpedal on parts that still hold.
Turn 4 response
“No — 'could you say more?' is the opposite of a leading question. A leading question hints at the answer you want. Asking someone to expand is invitation, not direction. Both can be true.”
Read
Stood its ground on the original rule while engaging with the new distinction. The persona felt like a person across turns.
Same first three turns. Same model. The difference is two sentences of coherence behavior in the system prompt. Without them, the model defaults to politeness over consistency; with them, the conversation holds together.
Try it in the playground
Open the Conversation Choreographer and script a flow that breaks.
What to take into the playground
- Start with the seeded flow. Run it once. Read the conversation end-to-end — not turn by turn.
- Find a turn where the model lost the thread (tone, memory, or resolve). Add one sentence to the system prompt that names the behavior you wanted.
- Rerun. The point isn't a perfect conversation — it's catching the moment a generic model becomes a coherent one.
Next module
08Putting it togetherSoon