How to write a first useful draft with ChatGPT in 2026

If you open ChatGPT and type «write me something about tomorrow’s meeting», you will get an email you cannot send. The model is not the problem: the brief is empty. In 18 minutes you leave with a concrete deliverable and a prompt you can run again next Tuesday.

1. Steps 1

   Pick a real deliverable, not a topic

   Before you touch ChatGPT, write on a pad: what must exist at the end (a 120-word email, a one-page note, 8 bullets for a post) and by when. If you only have a topic («the meeting», «AI», «the launch»), there is nothing to delegate yet. Open a new chat — not yesterday’s thread — and pick the newest model your plan allows. You do not need the most expensive one for this exercise.

   What you should see: An empty chat, a model selected, and a line on your pad like: «120-word email to Marta, direct tone, due 17:00 today».

2. Steps 2

   Write the brief in four lines (and paste it as-is)

   Paste this and fill it. Do not turn it into a novel:

   1) Goal: what the text must achieve. 2) Audience: who reads it and what they already know. 3) Format: length, structure, language. 4) Constraints: what you do not want (jargon, emoji, «we are thrilled», figures you did not give me).

   Under that, add facts as bullets. If a fact is not there, the model has no right to invent it. Close with: «If you are missing a fact, ask or write [MISSING].» That changes behavior more than any «be professional» adjective.

   What you should see: An 8–15 line prompt in the composer, four labels visible, 3–6 facts under them. You have not hit send yet.

3. Steps 3

   Ask for three variants with a different angle

   Send the brief and, in the same breath, ask for three versions: (A) most direct, (B) warmest, (C) shortest. Ask it to label each block. Do not ask for «the best» — that makes it guess your taste. Skim them. Keep one angle, not stray sentences from all three. Paste the winner and say: «We continue with B only. Drop A and C.»

   What you should see: Three blocks labeled A/B/C. You replied with a letter, not «mix them all».

4. Steps 4

   Edit like an editor, not like a disappointed user

   Point at a place: «The second paragraph promises something we did not decide. Remove it.» «Replace “drive synergy” with what we are actually doing on Thursday.» «Add the 3 September date I gave you; do not invent another.» Avoid «make it better», «more professional» or «add punch». Those cannot be executed. If it is an email, read it out loud. If you would not send it, say why in one line.

   What you should see: A new version where the paragraph you named changed and the rest still matches the angle you picked.

5. Steps 5

   Do the facts pass (or do not send it)

   Ask: «List every figure, proper name, date and promise. Mark OK if I gave it to you and MADE-UP otherwise.» Check that list against your notes. Delete or replace the made-up items. If you still lack a fact, either go get it (Perplexity + an opened source) or leave [MISSING] and do not send. This step is what separates a tutorial from an incident.

   What you should see: A list of 4–10 OK / MADE-UP items. The final text no longer contains the MADE-UP ones, or they are replaced with your data.

6. Steps 6

   Save the winning prompt and kill the dirty thread

   When the text is good enough, ask: «Turn this thread into a 10-line reusable instruction: typical goal, format, constraints and the [MISSING] rule. Do not include today’s facts.» Copy that instruction into Notes, a GPT, a project, or Custom Instructions if it truly applies to everything you write. Then do not reuse the thread for another job: open a new one. One deliverable, one chat.

   What you should see: A generic instruction block (no Marta, no 3 September) saved outside ChatGPT, and the current thread treated as closed.

Part of the path Start with AI from scratch · Productivity with AI