How to Prompt AI Like a Conversation, Not a Wish List

Why the first answer is a draft, and why polishing the first prompt wastes your time

Say you’re a sales operations manager, and a client’s procurement lead wants to know why an order is late. You open your AI tool and type, “Write an email to a client about a delayed order.” What comes back is polite, correct and no use to anyone. It apologises three times, promises a “swift resolution” and leaves out the new delivery date, which is the one thing she’ll look for.

So you do what most people do. You close the chat, write a longer prompt with everything you can think of in it, and paste that in. The second email is better but wrong in a fresh way: it’s so formal it reads like a legal notice. Twenty minutes in, you’re rewriting the prompt for the fourth time and nothing has been sent.

There was a faster route. Keep the first answer, tell the AI the one thing wrong with it (“Put the new delivery date in the first sentence and cut the apologies to one”), and read what comes back. That takes about a minute, and the second draft already carries everything the first one got right.

That’s iterative prompting: treating the exchange as a loop of draft, reaction and adjustment instead of a single request you have to get exactly right. The perfect first prompt is hard to write for a reason. Before you’ve seen a draft you often can’t say what you want, and a reaction like “too formal” only arrives once the formality is on the page.

None of this makes a careless opening prompt fine. A decent first prompt still saves turns, and we cover how to write one in our 4-part prompt formula and in the guide to reusable prompt patterns. This guide picks up where those stop, at the moment the first answer is on your screen and you have to decide what to say next.

The four-move loop: set the goal, find the gap, redirect once, confirm

The loop has four moves. The middle two repeat until the draft is good enough to use, and the first and last happen once.

Move 1: write a goal line before the first prompt

A goal line is a sentence about the outcome, not the document. For the sales operations manager it might read: “Email to a procurement lead explaining a two-week delay, so she keeps confidence in us and agrees to a Thursday call.” Paste it at the top of the first prompt and keep a copy, because you’ll measure every draft against it.

It takes about a minute to write.

If you can’t write the line yet, make the AI help. Add “Ask me up to three questions before you draft” to the first prompt. Those questions work like a cheap first draft, since each one shows you a decision you hadn’t made.

Move 2: name the gap by type

The most common reaction to a weak draft is a verdict: “make it better,” “try again,” “more professional.” A verdict gives the AI nothing to act on, so it changes things at random. Gaps come in a handful of types, and each has a phrasing that works. Here are the six we reach for most.

Move 3: change one thing and say which

When you change tone, length and structure in the same turn, you can’t tell which instruction caused the next problem. The AI also tends to over-correct everything at once, and you lose the parts of the draft that were working.

That’s our judgment from watching how drafts behave, so treat it as a default. Some gaps are bigger than one change, and the next section covers what to do then.

Move 4: confirm against the goal, then check the facts yourself

Once the draft feels right, ask the AI to read it against your goal line and say where it falls short. This catches things your eyes skip, like a missing ask. It doesn’t replace your own check on names, numbers and claims, which comes up again two sections on.

A worked exchange, start to finish, so you can see the loop run

Here’s the loop on a real-looking task. This is an illustrative exchange we wrote to show the shape, not a transcript from a client, and the replies are summarised, so your tool’s wording will differ.

Say you’re an account manager and a data migration for a client contact named Dana has slipped from October 14 to October 28 because her team’s export arrived with missing fields.

Four turns, and one email she could send.

Three things are worth noticing.

• The questions in turn 1 did real work. Whether to mention the cause at all was a decision the account manager hadn’t made, and she made it because she was asked.
• Each redirect carried one change. When the blaming line survived turn 3, she knew exactly which instruction hadn’t covered it.
• The confirm step caught a detail she hadn’t supplied. That “recovered most of the data” line would have gone to a client as a promise.

She still opens the project plan and checks both dates before she sends it. Confirming finds problems, but it doesn’t clear the draft to go out, and the next sections come back to that.

When the thread starts fighting you: drift, and when to start fresh

Say you’re a strategy consultant building a twelve-slide outline in one long chat over an afternoon. By the middle of the afternoon the AI has started reintroducing a section you cut early on, and an instruction to keep the language client-safe has quietly stopped applying. You keep typing “No, as I said earlier,” and each correction seems to make the next draft slightly worse.

That experience has some support from a preprint by Philippe Laban and colleagues. They ran simulated conversations with leading models and found an average drop of 39% across six generation tasks when a task was spread over several turns instead of given as one complete instruction. They traced much of it to unreliability rather than lost ability, and to models making assumptions early, then leaning on their own first attempts. In the authors’ words, “when LLMs take a wrong turn in a conversation, they get lost and do not recover.”^[1]

It’s a lab simulation on specific models and tasks, so it doesn’t tell you what your chat app will do on a Tuesday. What it does suggest is worth acting on. A correction layered on top of a bad draft inherits that draft’s structure, and your own earlier wrong turns stay in the thread as material for later replies.

That could sound like an argument against back-and-forth prompting. What it argues against is back-and-forth with no checkpoints, where nobody ever gathers up what’s been decided. The fix is a clean restart from a summary.

Two is our rule of thumb, not a measured threshold. Use it as a prompt to ask whether the thread has become the problem. Here’s the two-step restart, ready to paste.

The restart itself is cheap.

The two-sentence restatement in the new chat doubles as a drift test. If the AI’s version of your goal differs from yours, you’ve found the problem before it cost you a draft.

For the long-thread side of this, where answers degrade as the conversation grows, see our explainer on context rot. We won’t repeat it here.

When to stop prompting and go check: a five-question test

Say you’re a finance analyst working through a spreadsheet formula with AI. Six turns in, it looks right: it handles the totals, the dates and the grouping. Now you’re tempted to ask a seventh time whether it handles blank rows.

Another turn can’t settle that. Asking again can produce a more confident-sounding answer without producing a verified one. What settles it is pasting the formula into three rows you can compute by hand, one of them blank.

Three rows and a minute of work.

Iterating improves how well a draft fits what you asked for. In our judgment it does little to settle whether the draft’s claims are true. The model can match your goal line perfectly and still invent a date, a figure or a quote.

Use this five-question test before you send anything that has been through the loop. Score one point for each yes.

The right check depends on what the claim is. Recompute numbers, confirm names and dates against the original document, find quotes in the source they came from, and run formulas on cases you can work out by hand. Our fact-checking workflow goes through each of these, and the anti-hallucination toolkit covers prompts that reduce invented detail in the first place.

A score of 4 or 5 can also be a decision not to use AI for that piece. A reference letter, a legal clause and a sensitive message to an employee are all places where a person drafting is a reasonable answer.

How the loop changes with the size of the job and the role of the person running it

A single quick task and a three-hour piece of work both use the loop, but they don’t use it the same way. The same goes for the person at the keyboard, because what counts as “confirm” depends on what’s at stake in your job.

Where AI does real work in this loop, it helps to write down who owns what:

If you also want to see how a first prompt gets built before any of this starts, our piece on why prompting matters sets it out.

Making the loop a team habit instead of one person's trick

Start with one person’s habit and it stays one person’s habit.

One colleague who iterates well is a pleasant accident. A team that does it consistently needs something more deliberate, and the four-stage chain we use for rollouts applies here without changes.

Training builds the capability: people practise the loop on their own real tasks, not on a demo. Workflow design gives it somewhere to be used, for example a “goal line” field at the top of your team’s prompt template, and the restart prompt saved in your shared prompt library. Reinforcement is a manager asking in a review, “What was your goal line, and what did you change between the first draft and this one?” Adoption is what you see once those three have been running long enough that nobody needs reminding.

Measuring that goes beyond counting chats. Three questions follow our usual lens:

• Use: are people running multi-turn exchanges, or still restarting from scratch after each weak answer?
• Persistence: a month on, are goal lines still being written without a reminder?
• Impact: are first drafts reaching a sendable state in fewer rounds, and are reviewers fixing fewer invented details?

The right first step differs by role. A team lead owns the review question and the template field. An L&D lead owns the practice tasks for each function, and a head of finance or HR decides which outputs must pass the five-question test before they go anywhere.

A team can start with the tables in this guide and plenty do. Hands-on training shortens the learning curve in specific places: practising the loop on the team’s own tasks, agreeing what “confirm” means for each kind of output, and building shared habits instead of one enthusiastic individual’s. If that would help yours, our corporate AI training covers it.

Here’s a first step for this week. Take one email or summary you’d normally write from scratch, write the goal line first, and give the AI one change per turn until it’s right. Then run the five-question test before you send it.

Sana Mian, Co-Founder of Future Factors AI

Sana is an AI educator and learning designer specialising in making complex ideas stick for non-technical professionals. She has trained 2,000+ learners across corporate teams, bootcamps, and keynote stages. Future Factors offers AI Bootcamps, Corporate Workshops, and Speaking & Consulting for businesses ready to adopt AI without the overwhelm.

More about Sana →