Write prompts that get useful work (not vague answers)

What do you need before you start?

You need one real task, the facts only you know, and about 15 minutes to read the reply before you use it.

• One task you actually have today: a status email, a topic you are studying, or a weekend to plan
• The facts only you know: names, dates, limits, and what you already decided
• A free chat. ChatGPT, Gemini, and Claude all work

A prompt is the message you paste. ChatGPT for beginners covers the first account. This page is the next step: one prompt that finishes a real task.

How do you name the goal, the context, the output, and the boundaries?

Say what you want done, the facts only you know, the shape of the reply, and what must stay unchanged. Those four parts are the goal, the context, the output, and the boundaries.

Goal: [the one job, in one sentence].
Context: [the facts only you know]. Reader: [who this is for].
Output: [bullets, a short email, a table, or a list]. Length: [how long].
Boundaries: Use only the facts I typed. If a fact is missing, write MISSING. Do not invent names, dates, numbers, or sources. Do not do a second task.

Do this one task, then stop.

Fill the brackets before you send. If a line is still a bracket, the reply will guess.

How do you ask for one task only?

One prompt does one job. A status email is one job. A quiz is another. A weekend plan is another. Send the one that matches today.

Work. A status email your manager can read:

Goal: draft a status email I can send to my manager.
Context: Reader is [name, role]. This week: [what moved]. Stuck: [what slipped, and the real date]. I need: [the one ask, or "nothing"].
Output: an email under 120 words. Start with the result. Then what is stuck. Then what I need.
Boundaries: Use only the facts I typed. Do not add wins, numbers, or praise. If a date is missing, write MISSING. Do not send it. I will send it.

Do this one task, then stop.

School. A tutor quiz. You answer. The tool does not do the homework. Follow your school's AI policy.

Goal: tutor me on [topic] so I can explain it myself.
Context: Class is [class]. I already understand [what you know]. I am stuck on [what is hard]. My notes: [paste a few lines you wrote, or a short passage you are allowed to study].
Output: one question at a time. Wait for my answer. Then tell me what I got right, what I missed, and the next question.
Boundaries: This is for learning. Do not write anything I will turn in. Do not solve a graded problem. Do not write a paragraph I could paste into an assignment. Follow my teacher's rules. If I paste a graded question, refuse and ask me to try a practice question instead.

Do this one task, then stop.

Life. A weekend plan that stays inside your real limits:

Goal: plan Saturday and Sunday.
Context: Who is coming: [names]. Budget: [amount]. We must: [must-do]. We will not: [limits, including energy or travel time].
Output: a simple schedule with morning, afternoon, and evening. Then a short list of what I need to confirm.
Boundaries: Do not invent prices, hours, or reservations. Mark anything you are unsure about as CHECK THIS. Do not book anything.

Do this one task, then stop.

How do you change one thing and ask again?

Read the reply before you praise it. Say the one thing that is wrong, and ask for a new version of that part.

Change one thing only: [too long / wrong reader / a fact I did not type / missing the ask].
Keep everything else.
Use only the facts from my last message.
Do not add a new task.

If the reply did a second task you did not ask for, say so and tell it to drop that part.

How do you cut any fact you did not type?

A useful reply still needs a check. If you did not type it, it is not ready to use.

List every name, date, and number in your last reply.
Mark each one FROM ME or NOT FROM ME.
Delete the NOT FROM ME lines and show me the version I can use.
Do not put those lines back.

You still read that version. The label is a help. It is not the check.

What should you check before you use it?

Answer each line. If one is still a no, do not send, turn in, or book from the reply.

Director check: look before you use it

• Did the prompt name the goal, your facts, the output, and a boundary? If one part is missing, the reply is guessing.
• Was it one task? If the reply also wrote a second email, a plan, and a speech, cut back to the job you asked for.
• Can you point to where you typed each fact? If you cannot, delete that line.
• Is this schoolwork? You may use the quiz. You may not turn in the chat's words. Follow your school's AI policy.

What is one extra tip?

OpenAI's prompting guide says a stronger message names the goal, the context, the output, and the boundaries, then you shape the reply with a follow-up. Read Prompting and the help page on prompt engineering best practices. On this site, the same habit is Ask for a plan before the work when the job is big, and Stop generic AI answers when the reply could belong to anyone. The full set of stops is on Become an AI Director.