Iterate the conversation
~ min read
30-second summary
- The first answer is rarely the right one. People who use AI well keep the chat going instead of starting over.
- Three types of follow-up that solve almost everything: add a constraint, change the angle, deepen a piece.
- The AI remembers everything inside the same chat: you don’t have to repeat the context with every message.
- Better to start wide and narrow down with follow-ups than to pile thirty constraints into a single prompt.
- “Magic prompts” don’t exist. Iteration is the real trick.
You wrote a clear request, you pulled all the levers from the previous lesson, you hit enter. The answer is “ok”: not wrong, but not exactly what you needed. What do you do?
The most common reflex is to close the chat and try again with a new prompt, hoping to hit it right on the second try. It’s the wrong move. People who use AI well do the opposite: they keep the conversation going. An “ok” answer that becomes good with two lines of follow-up is the normal pattern, not the exception.
The iteration cycle
Section titled “The iteration cycle”Using AI well looks like a dialogue, not a command. You write the request, you read the answer, you say what to adjust, you get the better version. Repeat until you’re there.
Three types of follow-up cover ninety percent of the cases.
Add a constraint
Section titled “Add a constraint”Is the first answer too generic? Add a piece of context that narrows the field. “Budget max 50 euros”. “No dairy”. “For someone who’s never programmed”. You don’t rewrite the original request: you only add the new constraint.
Change the angle
Section titled “Change the angle”The answer is in the wrong shape, not in the wrong content. “Too long, give it to me in three bullets”. “Too technical, explain it like I’m ten”. “Skip the history, just the practical consequences today”. The role, the tone, the length: they change with one sentence.
Deepen a piece
Section titled “Deepen a piece”The overall answer is fine, but one part interests you more. “Explain point 2 better”. “Ok on the overall plan, but give me an hour-by-hour schedule for Saturday”. “Can you cite the two or three main sources?”. The AI doesn’t repeat everything: it develops only where you pointed the finger.
You can combine two of them in a single message (“make it shorter and add the 50-euro budget”) if the case is simple, or keep them separate across different turns if you want to see the effect of each constraint before adding the next.
One important thing: the AI remembers the whole chat. You don’t have to rewrite the initial context every time. If you told it earlier that you’re planning a weekend in Florence with a 200-euro budget, it keeps that in mind for all the later messages in the same conversation. You saw this in How does the AI “understand”?: there’s memory inside the same chat, none outside.
A new chat makes sense in two cases: when you’re actually switching topic (I stop planning the weekend, I start writing an email), and when the current chat has gotten stuck, the AI keeps making the same mistake, and two follow-ups aren’t enough. For everything else, stay on the same chat.
An example: planning a weekend
Section titled “An example: planning a weekend”The clearest way to see iteration in action is to watch it on a real case. Planning a weekend away is perfect: it starts generic and becomes yours through several turns.
Turn 1: the starting request
Section titled “Turn 1: the starting request”“Ok” answer, generic tourist guide. I start from here.
Turn 2: I add constraints
Section titled “Turn 2: I add constraints”Now the plan has my constraints. I didn’t rewrite the request from scratch: I only added what wasn’t there before.
Turn 3: I dig deeper
Section titled “Turn 3: I dig deeper”Turn 4: changing the angle
Section titled “Turn 4: changing the angle”Same information as before, different shape: same places, fewer stops, more room to breathe. It’s the third type of follow-up, “change the angle”: I don’t touch the content, I change how it reaches me.
The conversation now has a tailored answer. None of these four messages, by itself, would have been enough. Together, yes.
Practice with your AI
Section titled “Practice with your AI”Common mistakes
Section titled “Common mistakes”Three things that slow you down a lot while you’re learning to iterate.
Starting over instead of continuing. When an answer doesn’t work, the reflex is to close the chat and rewrite a “perfect” prompt. Often it’s slower. Adding one line to the existing chat is almost always faster and more precise: the AI keeps the work done so far.
Piling everything into a single request. The opposite temptation: putting thirty constraints into the first message to avoid later turns. It doesn’t work well. The AI loses pieces, forgets details, gets confused. A practical rule: if rereading the request there are more than two “and also”, you’ve probably loaded too much. Better to start wide and narrow down with follow-ups: the mistake shows up sooner, and it’s easier to fix.
Forgetting that the chat has memory. Repeating with every message “as I was saying” or “remember there are two of us” is wasted effort. Inside the same conversation the AI has everything in front of it. You only have to give context again when you open a new chat.
What comes next
Section titled “What comes next”Knowing how to ask and knowing how to iterate are the two abilities that hold up the rest of the manual. From here on, the lessons in the module are concrete use cases: understanding a complicated document, summarizing a long text, learning something new. In all these cases the two levers you’ve seen so far come back, tailored to the situation.