Plan, then write

What it is

Plan, then write splits one request into two. The first message asks for an outline and tells the model to stop there. You read it, move, cut or add sections, and send it back. The second message asks for the draft, built from the outline as you edited it.

The point is where your effort goes. A finished draft is persuasive: the prose is smooth, so a weak argument or a missing section is easy to miss. An outline shows the skeleton with nothing to hide behind. You can see in a minute that the risks come after the recommendation, or that the section on cost rests on one note.

It also changes what the model writes. With an agreed outline, every section has a job you approved, so the draft tends to stay on it instead of drifting into whatever the model thought a report of this kind should contain.

When it helps

Use it for long documents: reports, proposals, policies, business cases, training material. Use it for presentations, where the slide order is the argument. And use it when someone else has to agree the structure before you spend time on wording, such as a manager who wants to see the shape of a board paper first.

It is especially useful when you paste a lot of material. Ask the outline to say which notes support each section, and you can see before the draft exists whether the model has understood your material, missed a source, or built a section on almost nothing.

It also works when you know the content but not the order. Ask for two possible structures and pick one, as in ask for three versions.

When it does not

Short pieces do not need it. An outline for a five-sentence email is about as long as the email. Ask for the email and edit it.

Skip it if you will not read the outline. Current models already plan before they answer: Google’s guidance says Gemini’s recent models think internally, so it is generally not necessary to ask for the plan in the response. An outline only earns its extra turn when you review it. Printed just for the model’s benefit, it adds length and cost.

If the structure is fixed, give it rather than asking for one. When your company has a standard format for a business case, paste the headings and start the answer for it.

The draft can still drift. In a long chat, the model may follow its own first outline rather than your edits. Paste the agreed outline into the message that asks for the draft, and tell it not to add sections.

Before and after

Three documents from different kinds of work. The scores are the Prompt Grader’s free estimate with Written for set to Any chat model. A score counts what the prompt states; it does not grade the answer.

Write a report on why customers cancelled last quarter, using my notes below.

"""
[38 pages of exit-survey notes and call summaries]
"""

Exit-survey notes and call summaries from last quarter:
"""
[38 pages of notes]
"""

I need a report on why customers cancelled last quarter, for the leadership team, who will use it to decide where to spend next year's retention budget.

Do not write the report yet. First give me an outline: the section headings in order, one line under each saying what that section will argue, and which notes support it. Keep the outline under 200 words. If a section would rest on fewer than three notes, mark it [THIN]. A good outline lets me see the argument of the whole report in one minute.

The first prompt returns a long report whose structure you only see after reading all of it. The second returns a page you can check in a minute, with each section tied to its evidence and the weak ones marked [THIN] before they are dressed up in prose.

Make a presentation about our new expense policy for managers.

Expense policy:
"""
[expense policy text]
"""

I need a 10-minute presentation that walks our 30 people managers through the new expense policy above, so they can answer their teams' questions without calling finance.

Step 1, now: propose a numbered slide plan, no more than eight slides. For each slide, give a title that states the point in a full sentence and the two or three facts from the policy that back it. Leave out the history of the old policy. If the policy does not answer a question managers are likely to ask, list it at the end.

Step 2, after I approve the plan: write speaker notes of about 100 words a slide.

The bare request gets slide text written before anyone has agreed what the eight points are. The plan comes back as eight sentences you can reorder, plus a list of questions the policy leaves open, which you can settle with finance before the deck exists.

Write a remote work policy.

We are drafting a remote work policy for a 120-person accounting firm. Staff can work from home two days a week today, and the partners want that written down with clear rules for client meetings and equipment, so managers can answer requests the same way every time.

Before you draft, give me a numbered list of the sections the policy needs and the one decision each section needs from us, such as who approves a third home day.

Put [TBD] next to any decision you think we have not made yet. Do not draft any policy text until I confirm the list.

A policy written in one go fills every undecided rule with a sensible-sounding default. The plan turns those defaults into a short list of decisions, with the open ones marked [TBD], so the partners make them instead of the model.

The line to add

The first message asks for the plan and stops:

Do not write the {{document}} yet. First give me an outline: the headings in order, one line under each saying what that section will say, and which of my notes it draws on. Keep it under {{number}} words. I will edit the outline, then ask you to write from it.

The second message carries the outline you edited, so the draft follows your version and not the model’s first one:

Here is the outline we agreed. Write the {{document}} from it, in the same order, with no extra sections. If a section needs a fact that is not in my notes, write [CHECK] instead of filling it in.

{{agreed outline}}

For a long document, ask for one or two sections at a time. You read as it goes, and a problem in section two does not repeat through section eight. That is the same idea as splitting a big job.

In ChatGPT, Claude, Gemini and Copilot

The two messages work the same way in all four, because they only ask for text. Some research tools now build the step in. In the Gemini app, Deep Research creates a research plan from your prompt, and you can choose Edit plan before you start the research (Use Deep Research in Gemini Apps).

Google’s prompting guide adds the caution above: Gemini’s 2.5 and 3 series models generate internal thinking, so it is “generally not necessary” to have the model outline or plan in the response (Prompt design strategies). Read that as advice about reasoning. An outline you will edit is a different job, and it stays useful.

Anthropic’s guidance makes the same distinction for Claude. It says Claude handles most multistep reasoning internally, and that chaining prompts is still useful when you need to inspect intermediate outputs (Prompting best practices). Plan, then write is a two-step chain with you inspecting in the middle.

In any of the four, a long chat makes drift more likely. Paste the agreed outline into the second message rather than writing “use the outline above”.

Related

• Ask it to ask questions firstFill the gaps before the outline, so the plan rests on your facts.Pattern
• Split a big jobA chain of smaller prompts, each checked before the next.Pattern
• Draft, then critiqueCheck the finished draft against your criteria.Pattern
• Outline a presentation from your goalA slide-by-slide outline with one message per slide.Template
• Write a blog post from your outlineThe second step: a draft that keeps to your points.Template