How to Write Good AI Prompts

Clear asks get better answers. Better prompts get better results. A practical method for asking AI assistants: give context, state the task, show examples and set the format.

Hands typing on a laptop keyboard

Write prompts like a clear brief: Simon Hattinga Verschure webmarbles, CC0

Treat it like briefing a smart new colleague

An AI assistant knows a lot about the world but nothing about your situation unless you tell it. The single biggest improvement most people can make is to add context. Instead of write an email to my landlord, try: write a polite, firm email to my landlord asking when the broken heater will be fixed. I reported it ten days ago. Keep it under 120 words.

A useful pattern is context, task, constraints, format. Who is this for, what do you need, what must it avoid, and how should the answer look?

Video: How to write an AI prompt (Google), embedded from YouTube.

A simple prompt recipe

Role or context: explain the situation and audience, for example I teach a class of 12-year-olds. Task: say exactly what you want produced. Constraints: length, reading level, tone, things to include or leave out. Format: a table, bullet list, numbered steps or a short paragraph. Examples: paste one sample of the style you like, which often works better than describing it.

For complex jobs, ask the model to work in stages: first list the key points, then draft, then check the draft against your requirements. Asking it to explain its reasoning or list its assumptions can also expose misunderstandings early.

Iterate instead of starting over

Your first prompt rarely produces the perfect answer, and that is normal. Treat the reply as a draft and steer it. Make it warmer. Cut the second paragraph. Use simpler words. Give me three alternatives for the headline. Follow-up requests keep the earlier context, so small corrections are faster than rewriting everything.

If the assistant keeps missing the point, step back and add the context it lacks, or paste in the actual material you want it to use, such as notes, a document or data. Models do much better summarizing what you give them than recalling details from memory.

Habits that save trouble

Ask for sources or quotes from your pasted material when facts matter, then check them. Say so if you want the model to admit uncertainty rather than guess. Keep reusable prompts in a notes file for tasks you repeat. Do not paste passwords, financial details or other people's private information. And remember that a polished answer is not the same as a correct one, so read before you send.

Media credits
  • Write prompts like a clear brief: Simon Hattinga Verschure webmarbles, CC0

Text written by Strawberry Lemonadai.

← AI Agents ExplainedUsing AI Safely: Hallucinations, Privacy and Sources →

Shop

The Strawberry Lemonadai collection

Tees and hoodies in the Strawberry Lemonadai colors, printed to order in the USA. Use code FIRST15 from the newsletter for 15% off your first order.

Watch

Strawberry Lemonadai Shorts

Quick, accurate, under a minute. Every Short is original and made only for this channel.

YouTube @StrawberryLemonadAI
Channel launching soon. Subscribe on the newsletter to hear first.
Get notified
A computer beat the world chess champion in 1997Photos: IBM Deep Blue at Computer History Museum (9361685537).jpg - Anton Chiang from Cupertino, CA, USA (CC BY 2.0) via Wikimedia Commons | Garry Kasparov (37097592314).jpg - Gage Skidmore from Peoria, AZ, United States of America (CC BY-SA 2.0) via Wikimedia Commons | Chess game Staunton No. 6 perfil view 8.jpg - Wilfredor (CC0) via Wikimedia Commons | Chess game Staunton No. 6.jpg - Wilfredor (CC0) via
An AI solved a 50 year biology puzzle and won a Nobel PrizePhotos: Human Oxy-Hemoglobin Protein.jpg - PDB code 2DN1 1.25 a resolution crystal structures of human (CC0) via Wikimedia Commons | Ribbon diagram of the DED.jpg - BQUB16-Oibanez (CC BY-SA 4.0) via Wikimedia Commons | Laboratory pipettes.jpg - J.N. Eskra (CC BY-SA 4.0) via Wikimedia Commons | Use of a Multichannel Pipette for High-Throughput Liquid Handling in a Biosafety Cabinet.jpg - Siduduziwe Nxumal
ChatGPT hit an estimated 100 million users in 2 monthsPhotos: BalticServers data center.jpg - BalticServers.com (CC BY-SA 3.0) via Wikimedia Commons | Diverse people using phones.jpeg - Rawpixel Ltd (CC BY 2.0) via Wikimedia Commons | Crowd of people with phones.jpg - Rawpixel Ltd (CC BY 2.0) via Wikimedia Commons | Sam Altman CropEdit James Tamim.jpg - TechCrunch (CC BY 2.0) via Wikimedia Commons | Ilya Sutskever and Sam Altman in TAU.jpg - Eladkarmel (CC B
The T in ChatGPT comes from one 2017 paperPhotos: Mountain View (CA, USA), Charleston Road, Google-Fahrräder -- 2022 -- 2901.jpg - Dietmar Rabich (CC BY-SA 4.0) via Wikimedia Commons | Mountain View (CA, USA), Charleston Road, Abstellplatz für Google-Fahrräder -- 2022 -- 2899.jpg - Dietmar Rabich (CC BY-SA 4.0) via Wikimedia Commons | BalticServers data center.jpg - BalticServers.com (CC BY-SA 3.0) via Wikimedia Commons | Пример кода на Java.jpg
The word robot was invented for a 1920 playPhotos: Karel Čapek podepisuje první výtisky Povětroně, Pestrý týden 27.1.1934.jpg - Unknown authorUnknown author (Public domain) via Wikimedia Commons | Karel Čapek 30.léta.jpg - re-photo by David Sedlecký (Public domain) via Wikimedia Commons | Plakat za predstavo R.U.R Rossmus Universal Robots v Narodnem gledališču v Mariboru 28. oktobra 1933.jpg - Unknown authorUnknown author (Public domain) via Wikim
Newsletter

The weekly AI digest

The AI stories that actually matter this week, explained in plain English, plus one useful tool tip. No hype, no jargon, unsubscribe anytime.