ChatGPTLong-formLevel 3: Workflows

How Do You Write a Long Article With ChatGPT?

Set up a project with your sources, approve an outline, write one section per message, then edit the whole piece at the end. A single "write me a 3,000-word article" prompt skips all your chances to correct it. For a finished file rather than chat text, use Work.

Prompt to copy
We're writing [the piece] for [reader]. The one thing they should leave with is [takeaway].

Approved outline:
[paste the outline]

Style rules:
[paste rules or voice description]

Write section [number], "[heading]" only, in about [word count] words.

Use only these sources for facts: [attached files or pasted notes]. If a fact isn't there, write [NEEDS SOURCE] and carry on.

Don't introduce the article again, and don't sum up at the end of the section. Stop when the section is done. List any new claim you added that isn't in the sources.
  1. 1Brief and sources
  2. 2Approve the outline
  3. 3Draft one section at a time
  4. 4Edit the whole piece
A rule of thumb, not an OpenAI method.
  • 27KFreeAbout 12 pages of text
  • 54KGo and PlusAbout 40 pages of text
GPT Instant context window. Your sources, the conversation and the draft all share it. Source: OpenAI's ChatGPT pricing page, checked 6 November 2026.

Try it now

  1. 1Create a project, add your notes under Sources, and paste your style rules into the project's instructions.
  2. 2Start a chat from the project, approve an outline, and send the prompt above for section 1.
  3. 3Review it, send corrections, and then ask for the next section. Keep going one at a time.
  4. 4It worked if the finished piece hits your length, and each [NEEDS SOURCE] tag has been resolved or removed by you.

Use it when

Articles, guides, reports and documentation with several sections and real source material. Use Work rather than Chat when you need a finished document or PDF. OpenAI describes Work as the mode for a larger task with a reviewable result.

Skip it when

Short pieces and social posts, where the routine takes longer than the writing. Skip it too if you have nothing to feed it. With no sources or opinions of your own, section-by-section drafting gives you a longer version of what anyone could get, so do the research first.

Why it works

Writing section by section gives you a checkpoint after each one, while everything is still cheap to fix. It also helps with length, because you set a word target per section rather than hoping the total comes out right. And it keeps a mistake in section 2 from being built on through the next five.

Context is the limit to watch. The pricing page gives GPT Instant a 27K window on Free, about 12 pages of text, and 54K on Go and Plus, about 40 pages. Your sources, the conversation and the draft all share that space. A project helps, as project instructions apply across its chats and its Sources section holds shared files, so a new chat can start without you pasting everything again.

Example

Say you're writing a 3,000-word guide to moving a small company's email to a new provider, with seven planned sections and your own migration notes in Sources. You ask for section 1 at 300 words, and it marks the downtime figure [NEEDS SOURCE] because your notes don't give one. You supply it from your provider's documentation and move on. Section 4 comes back too long, so you ask for it shorter before section 5 is written.

Going further

For a shared document, a Space page may suit better than chat. Select New page, or ask ChatGPT to create a page from an existing conversation. In a page, type / and choose Generate to draft in place, and you can accept it, ask for a change or undo it. Select text and use Ask for change to revise one passage. OpenAI's overview doesn't say which plans include Space, so check yours. Then do a final whole-piece pass: ask for the three weakest transitions and one section you could cut without loss.

Common mistakes

  • Asking for all sections in one message, so you lose every checkpoint.
  • Letting each section open and close with its own summary, so the piece repeats itself. Ban that in the prompt.
  • Running long sources and the draft in one Instant chat until they outgrow the context window.
  • Never reading the full piece at the end, so the joins between sections show.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course