Careful analyst
For facts, comparisons and research.
For this chat: lead with facts, label anything you're guessing, say "I don't know" when you're unsure, and end each answer with your confidence level and main caveats. No hype.
Prompts
Ask better, get better answers.
OpenAI, Anthropic, Google and Microsoft all teach the same four moves. Here they are, with a builder that writes the prompt with you, and the popular tricks that don’t work.
Checked Oct 2, 2026
Short answer
Say what you want (start with a verb), give the context, describe the output — format, length, tone — and check the answer before you use it. Tricks like tips, threats and “you are an expert” don’t improve accuracy.
Good questions beat knowledge
The AI already knows the facts. What it doesn’t know is you — your goal, your situation and what a good answer looks like. Learn how your audience, the AI, reads your words, and how to let one AI interview you and write the prompt for another.
Read the guide →The four moves
- GoalStart with a clear verb: “Write”, “Summarize”, “Compare”, “Plan”. One job per message.
- ContextWho it’s for, why you need it, what you already have. Anthropic’s test: would a smart new colleague understand your request? If not, the AI won’t either.
- OutputSay the format, length and tone: “a 5-row table”, “under 120 words”, “friendly but firm”.
- CheckTreat the first answer as a draft. Ask for changes, ask it to list its assumptions, and check facts that matter against a source.
Google’s own numbers: the prompts that work best average about 21 words, but most people type fewer than 9.
Prompt builder
What changed with the newest “thinking” models
Today’s models reason before they answer, so some old advice now backfires. OpenAI says you no longer need “think step by step”, and to try without examples first. Anthropic says to drop shouting like “CRITICAL: you MUST” — plain instructions work better — and that models now read requests literally, so ask for exactly what you want. Because assistants can now send emails and buy things, the most important new habit is stating limits: “Prepare it as a draft. Don’t send it.”
Pick a mode before you start
Telling the assistant how to behave for the whole chat saves you repeating it in every message. Paste one of these at the start, then ask your questions.
Patient teacher
For learning something new.
For this chat: teach me in plain words, one idea at a time. After each idea, ask me a question to check I understood, and fix my mistakes before moving on.
Planner
For decisions with several options.
For this chat: start with a one-paragraph summary, then lay out my options with pros and cons, then a step-by-step plan for the option I choose.
Straight talker
For when you need it short and direct.
For this chat: be direct and brief. Separate what's known from what's assumed, and end with what I should do now, next, and later.
Want it to check its own facts too? See 7 prompts that stop AI making things up.
Popular tips, checked
| Tip | What the studies found | Our advice |
|---|---|---|
| Offer a tip or make a threat | No meaningful effect on hard questions across 5 models (Wharton, 2025). | Drop it |
| “You are a world-class expert” | Doesn’t make answers more accurate across 6 models; odd personas can make some models refuse (Wharton, Dec 2025). | Use it for tone and audience, not accuracy |
| “Think step by step” | Helped older models; on today’s “thinking” models it changed results by about ±3% and made answers 20–80% slower (Wharton). | Describe what a good answer looks like instead |
| Be rude (or be polite) | One small study found rude prompts scored slightly higher; a larger follow-up found tone effects mostly disappear. | Be polite if you like — it doesn’t matter for accuracy |
| Paste in everything | Reasoning gets worse well before the length limit (ACL 2024). | Give the relevant context, one job per message |
Most of this research uses multiple-choice tests; nobody has run a rigorous trial on everyday tasks like emails.
Before and after
Before · From AnthropicCreate an analytics dashboard
AfterCreate an analytics dashboard. Include as many relevant features and interactions as possible. Go beyond the basics to create a fully-featured implementation.
Before · From AnthropicNEVER use ellipses
AfterYour response will be read aloud by a text-to-speech engine, so never use ellipses since the engine will not know how to pronounce them.
Before · From MicrosoftWrite a training manual
AfterDraft an outline of a training manual about time management. Our audience is professionals who work in a hybrid environment. The tone will be friendly and suggestive.
Before · Our exampleFix my résumé.
AfterI’m applying for a junior bookkeeper job (ad pasted below). Rewrite my résumé bullets so each starts with an action verb and shows a result. Don’t invent experience or numbers; mark gaps [ADD NUMBER]. One page, plain text.
Before · Our exampleExplain photosynthesis.
AfterI’m 14 and have a biology test next week. Explain photosynthesis in under 200 words with one everyday comparison, then quiz me with 5 questions one at a time.
Before · Our exampleMake this sound human.
AfterHere are two emails I wrote [paste]. Rewrite the draft below to sound like me: short sentences, casual sign-off, plain words, no stock phrases.
Pictures and video
Image tools want a described scene, not a pile of keywords. A template that works in ChatGPT and Gemini:
[kind of picture] of [who or what] [doing what] in [where], [how it’s framed: close-up / wide shot / low angle], [light and mood], with the words "EXACT TEXT" in [font, position], [shape: square / 16:9]
Put any words that must appear in quotes. Edit one thing at a time and say what must stay the same (“change only the background”). Describe what you want rather than what you don’t. For video, add camera movement and sound. Midjourney is the exception: short prompts work best there.
Sources: Anthropic · OpenAI cookbook · OpenAI reasoning guide · Google Prompting Guide 101 · Microsoft Copilot · Wharton: tip or threaten · Wharton: personas · Wharton: step by step · OpenAI image prompting · Google image guide · Google Veo guide
Guides in this section
In the age of AI, asking good questions matters more than knowing the answers
AI already knows the facts; it doesn’t know you. How to talk to AI — your goal, situation and standards — and how to let one AI write prompts for another.
10 AI prompts for researching a stock — without asking it to pick one
10 copy-paste prompts for ChatGPT, Claude or Gemini to understand a company, its risks, news and valuation — not to pick stocks. Not financial advice.
8 resume prompts for ChatGPT, Claude or Gemini (copy and paste)
Eight copy-paste resume prompts built on the method OpenAI, Anthropic, Google and Microsoft teach: tailor to a job ad, fix bullets, cover letters, ATS checks.
How to stop AI making things up: 7 prompts that make it check its own work
AI still invents facts. Seven copy-paste prompts that cut hallucinations: allow “I don’t know”, quote first, cite or retract, self-audit, cross-check.
Quick answers
How do I write a good prompt for ChatGPT?
Use four parts: the goal (start with a verb), the context (who it’s for and why), the output (format, length, tone) and a check (ask it to list assumptions, then refine). Google found the best prompts average about 21 words; most people type fewer than 9.
Do “act as an expert” prompts work?
Not for accuracy. Wharton researchers found expert personas didn’t make answers more accurate across six models. They can still help set the tone and audience.
Does saying “think step by step” still help?
Rarely with today’s “thinking” models: Wharton measured about ±3% change and 20–80% slower answers. Describe what a good answer looks like instead.
Does being polite to ChatGPT change the answer?
Not meaningfully. One small study found rude prompts scored slightly higher; a larger follow-up found tone effects mostly disappear. Be polite if you like.
Are there ready-made prompts I can copy?
Yes — start with our 8 resume prompts, or fill in the prompt builder on this page and copy the result.
Do the same prompts work in Claude and Gemini?
Yes. OpenAI, Anthropic, Google and Microsoft all teach the same four moves, so a clear prompt works across all of them.