A short, hands-on field guide for the rest of us — the people using AI to write emails, plan trips, fix recipes, and make life slightly less chaotic. By the end you'll write prompts that actually work, every time.
Imagine asking a stranger to bake you a cake. If you say "make me food," they could come back with anything — toast, sushi, microwave popcorn. Now imagine giving them a recipe card with ingredients, steps, and an idea of what the finished cake should look like.
That's the difference between a vague AI request and a great one. The AI is the world's most eager, most well-read assistant — but it has zero clue what's in your head until you tell it. A prompt is how you put what's in your head onto a recipe card.
// The good news: you already know how to do this. You do it every time you brief a colleague, a designer, or a friend you're sending out for groceries. We're just going to make it slightly more deliberate.
Both of these prompts ask for an email. Only one will give you something you'd actually send. Read them side by side and you'll start to feel the pattern.
// The second prompt isn't "more technical." It's just specific where it matters: who the AI is, who you are, what changed, what you want, and what good looks like. That's what we'll learn next.
If your prompt has all five, you'll get a useful answer almost every time. Tap each card to see it explained with a real, daily-life example.
The background. Who you are, what's already happened, what matters.
Who you want the AI to be while it answers. A friendly therapist? A blunt editor? A patient teacher?
The actual task. One clear verb, ideally. Write, summarise, compare, list, rewrite, translate.
Length, tone, format, audience, what to avoid. The "no really, like this" details.
Show, don't just tell. One or two examples of what "good" looks like will out-perform 200 words of description, every time.
Steal these. Adapt the bracketed bits. Notice how every one has all five CRISP ingredients — that's the whole trick.
A colleague keeps roping you into "quick favours" that aren't quick.
[Role] Act as a kind, firm career coach. [Context] A peer keeps asking me to "quickly review" decks. Each one takes 45 minutes. It's eating my evenings. [Instruction] Write a short reply declining their next ask. [Specifics] Under 70 words. Warm. Don't apologise more than once. Offer one alternative (e.g. async comments by Friday). [Proof] Tone like: "Hey — would love to but…"
It's Sunday night and you have whatever you have.
[Role] Pretend you're a clever home cook who hates food waste. [Context] I have: half a cabbage, 3 eggs, leftover rice, soy sauce, one carrot, garlic, and a sad lemon. [Instruction] Suggest 2 dinner ideas using only these + pantry basics (oil, salt, pepper, chilli flakes). [Specifics] Under 25 minutes each. No new shopping. List ingredients first, then steps. [Proof] Format like a Notes-app recipe — no fluff, numbered steps.
You bookmarked it three weeks ago and still haven't read it.
[Role] Act as a thoughtful editor who briefs busy executives. [Context] I'm pasting an article below. I have 2 minutes. [Instruction] Give me: (1) the single core argument in one sentence, (2) 4 supporting points, (3) one question I should ask if I disagree. [Specifics] Plain English. No buzzwords. If the article is weak, say so. [Proof] Bullet style, like a daily briefing memo. [paste article here]
A friend asks "what should we do in Lisbon?" — you'd rather not Google for an hour.
[Role] Act as a local who hates touristy places but loves food. [Context] Two friends, 3 days in Lisbon, mid-budget, can walk a lot, will not wait in 90-min queues. [Instruction] Give us a day-by-day plan. [Specifics] Each day: morning + afternoon + 1 dinner spot. Mark anything that needs booking. Skip Belém Tower. [Proof] Style like a friend texting recommendations — informal, confident, one-liners.
You've tried a routine before and quit. Help yourself plan one you'll actually do.
[Role] Act as a behavioural-design coach who's allergic to hustle culture. [Context] I have 20 minutes a day, low energy in the morning, and I want to read more — say 4 books in 30 days. [Instruction] Design a realistic 30-day plan. [Specifics] Account for 4 "off" days. Suggest a tiny anchor habit. No 5-AM nonsense. End with one rule for when I miss a day. [Proof] Calendar-style table: Day | Action | Why.
You're spiralling about a choice. Make the AI argue both sides.
[Role] Act as two characters: an optimistic friend, and a sharp sceptic. [Context] I'm deciding whether to switch from a stable job to freelancing. I have 6 months of runway, a partner with steady income, and one client lined up. [Instruction] Each character writes me a 5-bullet pitch. [Specifics] Optimist: lead with the upside, then risks. Sceptic: lead with what could go wrong, then survivable upside. End with the 3 questions both agree I should answer first. [Proof] Format: ## Optimist / ## Sceptic / ## Both agree.
Most "the AI is dumb" moments are actually one of these. The fix is almost always a sentence or two longer.
"Best" is meaningless until you say best for what. Best cheap? Best beginner-friendly? Best for a vegan triathlete?
"Summarise this, then translate it, then make a poem out of it, then…" Split it. One verb per prompt is your friend.
"Write about climate change." For whom — a 9-year-old, a CEO, your aunt who reads the Daily Mail? They each need a different answer.
"Could you maybe sort of summarise this if it's not too much trouble?" The AI doesn't need feelings. It needs nouns and verbs.
The first answer is rarely the best. Reply with: "shorter," "more skeptical," "rewrite as a list," "now in my voice." Iterate, don't restart.
"No emojis, no hashtags, no exclamation marks" saves more time than 200 words of positive instructions.
Models can confidently invent statistics. For anything factual, say: "If you're not sure, say so." Then verify.
If you can paste one sample of what good looks like, do it. One example is worth a paragraph of description.
Your friend wants help writing a wedding toast. Drag the right tiles into the prompt tray, in any order. Watch out — three of the tiles don't belong. Hit "Check" when you're done.
Six quick questions. No trick answers — just situations you'll genuinely run into. You'll get an explanation after each one.
Paste any prompt — real or made up. We'll check it for the five CRISP ingredients and tell you what's missing. Nothing leaves your browser; this is rule-based, not a model call.
If you remember nothing else, remember CRISP. Print this, screenshot it, tape it to the side of your monitor. You're done with the handbook.
"I'm a __ trying to __. Last week, __."
"Act as a __ who __."
"Write / list / summarise / compare __."
Length, tone, format, audience, what to avoid.
"Examples of the vibe: __, __. Avoid: __."
Want more like this?
More handbooks on AI, engineering, and building in public — free to read, always open source.
Browse all handbooks →