Guide · Prompts that work
The Prompt Playbook
A prompt is a set of instructions. Better instructions get better answers — not because of magic words, but because the AI finally knows what you actually want. This guide teaches one framework that works on every task, with real before-and-after examples.
No "secret prompts" here. There is no perfect prompt that works everywhere — prompt quality depends on the task. This guide teaches a transferable skill, not a collection to memorize.
The problem
Vague in, vague out
Imagine asking a coworker: "Write something about the meeting." They'd stare at you. Write what — an agenda? A summary? A complaint? How long? For whom? You'd clarify, because humans know to ask. But with AI, most people hit enter on the first draft of their thought and accept whatever comes back.
Here's that same habit in the wild:
"Write an email to my landlord."
A stiff, generic template — wrong tone, wrong situation, wrong details. Now you're editing someone else's guess instead of writing your own email.
The AI didn't fail you. It did exactly what you asked — which was almost nothing. Every unanswered question in your prompt is a guess the AI has to make on your behalf. The framework below is just a checklist for closing those questions before you hit enter.
The framework
Four pieces, every time
Good prompts are built from the same four parts. You don't need all four for every task — but checking the list is the skill.
GOAL
What are you trying to accomplish?
The one-sentence job. Without a goal, the AI optimizes for "sounds plausible" — which is how you get confident-sounding fluff. State the job and the answer bends toward it.
CONTEXT
What background changes the answer?
Who it's for, what it builds on, what's already true. Context is the difference between a generic answer and yours. "For my 10th-grade class" reshapes every sentence that follows.
CONSTRAINTS
What's off-limits or fixed?
Length, tone, budget, "don't do this." Constraints don't limit the AI — they aim it. "Under 150 words" and "no jargon" remove entire categories of wrong answers.
OUTPUT
What should the answer look like?
Format, structure, what to skip. Specifying the shape of the answer saves you the re-formatting pass. "A numbered list, one line each, no intro paragraph" lands ready to use.
An AI answers by predicting what should come next. Every goal, detail, constraint, and format choice you give removes wrong predictions from its options — you're not "unlocking" hidden abilities, you're steering a fast guesser with better directions. And remember: the mix changes with the task. A coding question needs more context; a brainstorm needs more constraints. There is no single perfect prompt — only the right ingredients for the job.
Worked examples
Six prompts, repaired
Each one shows the bad version, the better version, and — the actual lesson — why the changes help.
School — the essay
"Write an essay about climate change for my history class."
"I'm writing a 500-word history essay for my 10th-grade class. Goal: explain how industrialization accelerated climate change, in three body paragraphs. Use plain language a 15-year-old can follow, point to the kinds of sources I'd be expected to use (books and journals, not blogs), and end with two study questions I can quiz myself on."
The bad prompt outsources the whole assignment and teaches nothing. The better one states the goal (explain a specific mechanism), the context (10th-grade history), the constraints (500 words, three paragraphs, serious sources), and the output (draft plus self-test questions). An essay-shaped answer plus study questions supports the assignment instead of replacing it.
Email — the raise request
"Write an email asking for a raise."
"Draft a 150-word email to my manager asking for a 5% raise. Context: I've been on the team two years and just led the inventory migration that finished three weeks early. Tone: confident but respectful — no begging, no threats. Output: just the email body with a subject line."
"Asking for a raise" hides everything the AI needs: how much, why you deserve it, who reads it, how long it should be. Naming the constraint (150 words, a specific tone) and the context (the migration win) turns a generic template into a draft you could actually send — and the output spec ("just the body") stops the AI padding your email with commentary about emails.
Research — the health question
"Tell me about keto."
"I'm considering a keto diet and want to understand the tradeoffs before talking to my doctor. Goal: a balanced overview of what keto involves, who it tends to suit, and common side effects. Keep it to about 300 words, no medical advice aimed at me personally, and list three reputable sources I can read myself."
"Tell me about keto" has no goal, so the AI picks one for you — and on health topics, a wrong guess has real stakes. The better prompt states the goal (an overview to prepare for a doctor conversation), the constraint (300 words, no personal medical advice), and the output (balanced summary plus sources). The transferable habit: the more the answer matters, the more precisely you should frame the question.
Coding — the bug
"My JavaScript doesn't work. Fix it."
"This function should filter a list of tasks by due date, but it returns an empty array even when every task has one. [paste the code here]. Goal: explain the bug in one or two sentences, then show the smallest possible fix. In the explanation, use plain language so I understand the logic — not just the code."
An AI can't fix code it can't see — the bad prompt forgets the single most important piece of context: the code itself. The better prompt adds the expected-vs-actual behavior ("returns empty" points straight at the bug), constrains the answer (explain briefly, smallest fix), and asks for the reasoning. That last line matters: prompts that ask for the why make you better at the next bug, too.
Planning — the weekend trip
"Plan a weekend trip to Philadelphia."
"Plan a 2-day weekend trip to Philadelphia for 2 adults and a newborn. Goal: keep it low-key — one or two things per day, near public transit. Budget: under $300 total excluding the hotel. Output: a day-by-day plan with times, each activity tagged stroller-friendly or not, plus a short packing note for the baby."
"Plan a weekend trip" is really ten questions disguised as one — who, what budget, what pace, what matters. Every extra detail is load-bearing: two adults plus a newborn changes the pace, the budget changes the activities, and the output format (day-by-day with tags) means the answer is usable without re-sorting it yourself. Constraints don't limit the AI; they point it.
Brainstorming — the business name
"Give me business name ideas."
"Brainstorm 10 name ideas for a mobile dog-grooming business in Cape Coral, Florida. Goal: friendly and trustworthy, not corporate. Avoid paw puns — too common here. For each name, add one sentence on the vibe it gives. Output: just the numbered list, no intro paragraph."
Without constraints, "name ideas" produces the same generic list everyone gets. The better prompt gives the context (mobile grooming, one specific city), an anti-example ("avoid paw puns" is a constraint too), and the output format (numbered list, one vibe sentence each, no intro). "No intro paragraph" removes the filler nobody asked for — specifying what you don't want is often the fastest improvement.
Watch out
Three common traps
These show up in almost everyone's prompts. Spot them in yours and fix them first.
Irrelevant detail
The long backstory about your day that doesn't change the answer. Every extra sentence is a chance for the AI to latch onto the wrong thing.
"It was raining Tuesday, my dog ate my homework, and by the way — write an essay about climate change."
Fix: keep only the facts that change the answer. Trim the rest.
Contradictory instructions
"Keep it short but cover everything in depth, casual but professional." Instructions that cancel each other out force the AI to guess which one you meant — and its guess won't be yours.
"Summarize this in 50 words with full detail and nuance."
Fix: pick one lane per axis. If you truly need both, split the task into two passes.
Everything at once
"Plan my trip, write the invite, and draft the budget." Three jobs, one prompt, three mediocre answers. Bundled prompts get bundled thinking.
"Also redesign the logo and explain quantum computing while you're at it."
Fix: split the jobs. Use the output of one as the context for the next — chains beat bundles.
Try it
Practice in the Prompt Repair game
Reading the framework is step one. Step two is reps: in Prompt Repair you get broken prompts and fix them yourself — pick the missing piece, see the answer improve, and build the habit the way it actually sticks: by doing.
Repair prompts yourself in the gameThe email guide
Send me the Prompt Playbook
The full playbook as a printable one-page reference: the four-part framework, all six repaired prompts, and the trap checklist — so it's in your inbox next time a prompt goes sideways.
Keep learning
Where this goes next
Prompts are how you talk to AI. If you're curious how the talking actually works underneath, How an LLM Works shows what happens between your prompt and the answer. Or head back to the Learn hub for the rest of the guides.