How do you have a back-and-forth conversation with ChatGPT?

OpenAI released ChatGPT on Wednesday as a free research preview. What sets it apart from earlier GPT-3 tools is the dialogue format. Instead of writing one carefully engineered prompt and hoping for the best, you can talk: ask something, read the answer, and reply with "shorter," "make it cheaper" or "what about the second idea?" OpenAI's launch post says this format lets ChatGPT answer follow-up questions, admit mistakes, push back on wrong assumptions and decline inappropriate requests.

That changes how you get good results. With a one-shot prompt, all the effort goes into the first message. In a conversation, the first message can be simple, and the quality comes from the steps that follow.

Why follow-ups work

One of OpenAI's own sample chats shows the idea. A user asks what Fermat's little theorem is, then asks "how is it used in cryptography?" without repeating the theorem's name. ChatGPT understands that "it" points back to the earlier question. Another sample shows a user giving follow-up instructions, such as asking it to summarize the conversation so far.

This means you can work in small moves:

• Correct: "No, the party is indoors."
• Narrow: "Only the ideas under $50."
• Reshape: "Turn that into a checklist."
• Probe: "Why did you suggest that one?"

Each reply gives ChatGPT more information than any single prompt would, and you only add what you discover you need.

A worked example: planning a birthday party

Here is a five-turn conversation. The replies are summarized and illustrative; your results will differ.

Turn 1: start broad.

I'm planning a birthday party for my daughter, who is turning 8. Can you suggest some themes?

ChatGPT lists several themes, such as space, a science lab, a treasure hunt and a garden tea party, with a line about each.

Turn 2: add the constraints you forgot.

We'll have 12 kids, it's in our apartment, and my budget is about $150. Which of those themes work best?

It drops the ones that need a yard and recommends the science lab and treasure hunt, explaining why they suit a small indoor space.

Turn 3: pick one and ask for structure.

Let's go with the science lab. Give me a schedule for a two-hour party.

It produces a timeline: arrival and lab coats, two experiments, snack break, cake, a goody-bag station.

Turn 4: correct a detail.

Two of the kids have nut allergies, and I'd rather avoid anything that stains. Update the plan.

It swaps out the messier experiment and suggests nut-free snacks. You did not have to restate the theme, the age, the budget or the schedule.

Turn 5: change the format.

Now turn the whole plan into a shopping list grouped by store section, with rough prices.

The final answer combines everything from the earlier turns into one practical list. Check the prices yourself: ChatGPT is not looking at any store's website, so its numbers are guesses.

Habits that help

Give one change at a time when you can, so it is easy to see what each instruction did. If an answer goes off track, say what was wrong rather than starting over: "That's too formal; write it the way you'd talk to a neighbor." Ask it to explain or show its reasoning when a recommendation surprises you.

Keep the limits in mind. OpenAI warns that ChatGPT can write answers that sound plausible but are wrong, and The Verge found it confidently presenting made-up details as fact. Its knowledge of events after 2021 is limited. It also tends to guess what you meant when a question is ambiguous, rather than asking. A follow-up such as "What do you need to know to answer this well?" invites it to ask you questions instead. And as with any service still in testing, it is sensible to keep private information out of your chats.

How does it keep track of earlier messages in a chat?

ChatGPT is built on OpenAI's GPT-3.5 series of language models, and a language model does not have a memory in the everyday sense. It reads a stretch of text and predicts what should come next. OpenAI has not published the exact details of how ChatGPT assembles a conversation, but the basic principle for models in this family is that the earlier turns of a chat are part of what the model reads when it writes its next reply. Your messages and its answers form a running transcript, and that transcript is the context for each new response.

That explains both the strengths and the limits. Because earlier turns are in view, a word like "it" or "the second idea" makes sense. But the amount of text a model can read at once, its context window, is finite and measured in tokens, the word pieces these models work with. For the standard GPT-3 models, OpenAI documents a limit of 2,048 tokens shared between the prompt and the reply, roughly 1,500 words. ChatGPT's own limit has not been published, but it is reasonable to expect that in a very long chat the oldest details may fall out of view or get less attention. If something early matters, repeat it, or ask for a summary and continue from that.

It also means the model does not learn from you over time. What it "remembers" lives in the current conversation only. Start a new chat and it begins with a blank page, so paste in any background it needs again.