Use it like an editor, not a vending machine

Writing with ChatGPT works when you keep the judgment and outsource the scaffolding. The people who get generic sludge are usually asking for the finished thing in one move. The people who get useful drafts are running a process: brief, outline, draft in sections, then an edit pass with rules.

This is the writing workflow used across the chatpt.chat field guide, and it works the same whether you draft in ChatGPT or anywhere else.

This is that process. It takes about the same time as writing badly on your own, and it produces something you can publish.

Stage 1 — Write the brief before you open the chat

A writing brief has five fields. Fill them in a note first; paste the note into the prompt. It takes two minutes and saves three drafts.

• Reader. Who is this for, and what do they already believe? “Operations managers who have ignored two vendor emails” is a reader. “Business professionals” is not.
• Purpose. What should change after reading? A decision, a reply, a click, a calmer mood?
• Shape. Word count, format, and where it will live: email, doc, landing page, slide notes, Slack.
• Voice. Two adjectives and one sample. “Dry, specific. Here is a paragraph I wrote last month that sounds like me: …”
• Fixed facts. The numbers, names, dates, and claims that must appear exactly. Mark anything the model must not invent.

The voice sample is the most useful line in the brief. A 60-word sample of your own writing teaches more about register than any adjective.

Stage 2 — Ask for outlines, not prose

Three outlines with different arguments beat one outline with three sections. You are shopping for a structure, and structure is where most writing fails.

Here is my brief: [paste brief].

Give me three outlines for this piece. Each outline must argue something different, not
just reorder the same points.

For each: the one-sentence argument, the section headings in order, and the evidence each
section needs.

Then pick one and say why. The model will follow your reasoning in the next step. If all three outlines are the same, your brief has only one idea in it — go back to stage 1.

Stage 3 — Draft in sections, not in one shot

A 1,500-word generation is a gamble. A 250-word section with a clear job is a draft you can steer.

Write section 2 only: "Why the current process breaks at handoff".

Rules:
- 220-260 words.
- Start with the concrete failure, not a definition.
- Use the two details from my brief. Do not add new examples.
- End with a sentence that makes section 3 necessary.

Then read it before generating the next one. If section 2 drifted, fix it now; otherwise the drift compounds.

Two habits that help:

• Give the previous section as context. “Here is what section 1 says. Do not repeat it.”
• Ask for one deliberate risk per section. A specific opinion, an unusual example, a blunt sentence. Average prose is safe prose; you want one thing per section that a person would actually say.

Stage 4 — Edit with rules, not vibes

“Make it better” produces a different draft, not a better one. Use an edit pass with explicit rules.

Edit the draft below. Keep my voice and my argument.

Rules:
- Cut filler, throat-clearing, and any sentence that restates the previous one.
- Do not add facts, examples, statistics, or quotes.
- Replace abstractions with the concrete details already in the draft.
- Keep the paragraph order.
- Flag any sentence that sounds like it could appear in any article on this topic.

Return the edited draft, then a list of the five biggest cuts and why.

That last instruction — flag the generic sentences — is the one that changes the output. Models are good at spotting their own register when you ask directly.

Then do a human pass for three things a model cannot do for you:

1. Facts. Every number, name, and quote gets checked against a source.
2. Opinions. If you would not say it out loud to a colleague, cut it.
3. Risk. Claims about people, companies, health, money, or law get a second reader.

Worked example: the client update

Brief. Reader: a client who is anxious about a delay. Purpose: keep trust, no over-promising. Shape: 150-word email, subject line. Voice: plain, calm, no corporate softening. Fixed facts: new date is 12 November; the delay came from the payment provider’s review, not our build.

Prompt.

You are a project lead writing to a client. Write a 150-word email about a two-week delay.

Facts you must use exactly: new launch date is 12 November. The delay came from the
payment provider's review, not from our build. We have already completed the integration
and are using the wait to test refunds.

Constraints: no apology theatre, no "we hear you", no passive voice hiding who is doing
what. One clear next step with a date. Subject line under 8 words.

What to expect. A usable draft in one pass, because the brief contains the facts and the emotional job. The edit pass then cuts the sentence that says “we understand this may be frustrating” — unless you actually want it.

Worked example: the newsletter intro

The failure mode here is a 300-word warm-up that says “AI is changing everything”.

Write the opening of a weekly newsletter for independent bike shop owners.

The issue is about cash flow in the slow season.

Rules:
- 90 words maximum.
- Start with a specific scene: a shop owner counting a drawer on a Tuesday.
- No mention of "the industry", "the landscape", or "these uncertain times".
- End on the practical question the issue answers.

Specific scene, hard word cap, banned phrases. That is a brief. The result will still need your voice, but it will not need a rewrite.

The generic-voice trap

If the draft reads like it was written by a committee, one of four things is missing:

You can also ban the tells directly: “no ‘in today’s fast-paced world’, no ‘unlock’, no ‘delve’, no rule-of-three lists, no em-dash pileups.” It feels petty. It works.

What not to outsource

• Facts and quotes. A fabricated quote in a published piece is a reputational event, not a typo.
• Your opinion. If the piece argues something you do not believe, you will not defend it when someone pushes back.
• Personal experience. “When I worked at X” cannot be generated. Either it happened or it did not.
• Disclosure. If a publication requires you to say how AI was used, do it. Policies differ by outlet, client, and country; check the one that applies to you.
• The final read. Read it aloud. Anything you stumble on gets cut or rewritten.

If the writing is a job application, the rule is stricter: a model may change how a true thing is phrased, never whether it is true. The job-hunt guide has the evidence-first workflow.

A compact checklist

Before you send anything:

• ☐ The first sentence contains the point, not a wind-up.
• ☐ Every number traces to a source you can open.
• ☐ No sentence survives only because it sounds nice.
• ☐ The reader is named in the brief and addressed in the text.
• ☐ One opinion is clearly yours.
• ☐ Voice sample was used, not just adjectives.
• ☐ Read aloud once.

Related

• A prompt is a brief, not a wish — the five parts of a prompt, with more before-and-afters.
• The office workflows worth automating first — email, meeting notes, and status updates.
• Prompts for writing — 17 prompts that follow this workflow.
• Open the workbench — turn a one-line request into a full brief.