“Write a report on customer satisfaction.”
A practical guide to AI prompts
Write better prompts.
Get better answers.
Learn how to write an effective prompt for ChatGPT, Claude and Gemini: define the goal, supply relevant context, set real constraints and make the answer possible to verify.
Free · no sign-up · the tools run in your browser
A good brief5 parts
01GoalWhat actually has to be finished?
02ContextWhat can the model not know?
03ConstraintsWhat must be included, and what is out of scope?
04FormatHow will the answer be used?
05VerificationWhat does a good answer look like?
The short version
What is an AI prompt?
A prompt is the message or instruction you give an AI tool. An effective prompt does more than name a topic: it explains what to produce, who it is for, what context the model needs, which constraints matter and what the answer should look like.
You do not need a magic formula. Start with five parts: goal, context, constraints, format and verification. Use only the parts that genuinely affect the task.
Example of an effective prompt
Write an email of no more than 120 words to a long-standing customer waiting for a delivery.
Context: the delivery has moved from 28 to 31 July because of our warehouse error.
Constraints: apologise specifically without blaming the carrier. Confirm the new date. Do not invent details or offer a discount.
Format: subject line and two short paragraphs.The example works because it removes important guesses, not because it is long.
See the difference
Same task. Completely different brief.
Switch between vague and effective use. The example shows why a polite answer is not the same as a usable one.
“Customer satisfaction is important for every business. Companies should listen to their customers and continuously improve their services …”
32 / 100 usability
The model has no data, no audience, no decision to support and no standard of evidence.
Open the full answer lab →Start here
Stop writing longer. Start briefing better.
A prompt is not an incantation. It is a brief. Include what changes the decisions the model has to make, and cut the decoration that does not.
Define the deliverable
Replace “help me with marketing” with a specific output, a named audience and the effect you want.
Add only decision-relevant context
Tell the model what it needs in order to choose correctly. Not your entire back story.
Make quality visible
State the format, the constraints, the sources or the checks that let you judge the answer.
Working tool
Build your first draft here
Answer six short questions. The prompt is assembled locally on your device.
Your prompt
Fill in the fields and your prompt appears here.
Nothing you type is sent to us.
Pick your task
Guides for what you are actually doing
Writing and editing
Why AI prose sounds like AI prose, at the mechanism level, and how to make it stop.
Read the guide →Research and fact-checking
Separate findings, sources, inference and uncertainty so the result can be checked.
Read the guide →Code and agents
Fence the scope, define done as something testable, and demand the real test output.
Read the guide →Image prompts
Subject, composition, light and style, plus when the result has to be labelled.
Read the guide →Analysis and data
What models are reliable at, and the arithmetic you should never let them do silently.
Read the guide →Work and productivity
Meeting notes, decision memos and plans, without invented owners and deadlines.
Read the guide →Going deeper
The prompt is only one part of the system
Modern AI work is also about context, tools, security, regulation and knowing which advice has quietly expired. Each of these pages carries the date it was last verified.
Why the AI ignores your instructions
The most common complaint there is, explained by mechanism rather than by folklore.
Read the guide →Prompting reasoning models
Reasoning control moved into an API parameter. The cross-vendor table, and what to stop writing.
Read the guide →Connectors, MCP and agent tools
What connecting an assistant to your files actually does, in plain English.
Read the guide →The hidden cost of structured output
Forcing JSON is not free. Four 2026 studies, and no vendor documentation mentions it.
Read the guide →Context engineering
Why quality degrades before the context window fills, and what to do about it.
Read the guide →AI and the rules
What the EU AI Act actually requires of ordinary businesses using AI.
Read the guide →Tested, not repeated
Some of what everyone “knows” turns out to be wrong
Most prompting advice has never been tested. Some of it has, and did not survive. Where nobody has looked, we say that instead of picking a side.
Prompt folklore, tested
Tipping, threats, emotional appeals and the “you are an expert” opener that almost every guide recommends.
Read the guide →AI detectors and false accusations
Why the false positives land hardest on non-native English writers, and what to do if it happens to you.
Read the guide →Verify before you cite
A fabricated citation went from a Reddit post to a judge's signed order. Nobody opened it.
Read the guide →Most important of all
Test the prompt. Do not fall in love with it.
A prompt is good when it produces good results on representative tasks. Not when it looks sophisticated.
Build a simple evaluation routineBefore you publish
- Are the claims sourced?
- Are the assumptions visible?
- Is the format actually usable?
- Does it hold up on hard examples?
- Has sensitive data been removed?