Give the model a job, not a topic
A useful request names the goal, the relevant context, the constraints, and the shape of the answer. More words do not automatically make a better prompt.
The four layers of a useful request
Build the request from the outcome upward. Each layer removes a different kind of ambiguity.
- GoalThe decision or artifact you need
- ContextOnly the facts that change the answer
- ConstraintsAudience, limits, tools, and deadline
- OutputA format you can inspect and use
- Goalground it: Context
- Contextbound it: Constraints
- Constraintsshape it: Output
- Goal
Say what decision or artifact you need.
- Context
Include only facts that change the answer.
- Constraints
State limits such as audience, length, tools, or deadline.
- Output
Ask for a format you can inspect and use.
Tell me about AI security.
Create a five-item checklist for a beginner reviewing whether an AI chat contains private customer data. Explain each check in one sentence.
Choose details that change the answer
Useful context is selective. A deadline changes a project plan; the color of your desk does not. Before adding a detail, ask whether a different value would produce a different answer. If not, leave it out. This keeps the request easier to read and reduces the chance of sharing information the model never needed.
| Part | Question to answer | Example |
|---|---|---|
| Goal | What useful result do I need? | A launch checklist |
| Context | Which facts change that result? | Solo developer; web app is already built |
| Constraints | What limits must the answer respect? | Five steps; no paid tools; include privacy |
| Output | How should the result be organized? | A table with action, reason, and check |
Build the request in four passes
From a rough idea to an inspectable request
Do not search for one magical sentence. Improve the request one decision at a time.
- Rough needName the real task
- Bounded taskAdd audience and limits
- Safe contextRemove secrets and identifiers
- Inspectable outputChoose a useful format
- Rough needclarify: Bounded task
- Bounded taskprotect: Safe context
- Safe contextshape: Inspectable output
- Write the result first
Start with a verb such as compare, draft, diagnose, or plan.
- Add the user and situation
State who will use the result and what has already happened.
- Set boundaries
Limit scope, length, sources, tools, cost, or tone when those limits matter.
- Define a checkable format
Ask for steps, a table, a checklist, or another structure you can inspect.
Run a privacy filter before you send
Treat the prompt box like an external form. Information should enter only when it is necessary and permitted. Names can become roles, real numbers can become representative ranges, and credentials should never be pasted. If the task cannot be explained safely, stop and use an approved internal process instead.
Send, replace, or stop?
Classify each piece of context before it reaches the AI tool.
- Does it change the answer?If no, remove it
- Is it safe and permitted?Public or approved context may continue
- Can it be anonymized?Use placeholders or ranges
- Sensitive and necessaryStop and use an approved workflow
- Does it change the answer?yes: Is it safe and permitted?
- Is it safe and permitted?not as written: Can it be anonymized?
- Can it be anonymized?cannot sanitize: Sensitive and necessary
Work through a complete example
Help me launch my app. Here are my customer list and production credentials.
Create a five-step launch checklist for a solo developer releasing a small web app. Include accessibility, privacy, rollback, and one verification check per step. Use a table. Do not assume paid tools.
The improved request removes private data, identifies the user, names the artifact, sets the scope, and asks for evidence. It still does not guarantee a correct answer. It simply produces an answer that is easier to inspect—the skill you will practice in the next lesson.
Practice with a reusable prompt card
| Prompt field | Your note | Ready when… |
|---|---|---|
| Result | What should exist when the task is done? | It begins with a clear action verb |
| Context | What does the assistant truly need to know? | Every detail can change the answer |
| Limits | What must it avoid or respect? | Risk, audience, time, and tools are bounded |
| Format | How will you inspect the result? | The output can be checked item by item |
| Privacy | What must be removed or replaced? | No secrets or unnecessary personal data remain |
Check your understanding
Three practice questions. They are not saved to an account or awarded points.