Asking for what you want

Start with a real job

You have a page of notes about a science experiment and want help organising them. “Make this good” gives the system very little direction. We will build a prompt one decision at a time, then inspect the response.

Use invented or appropriate non-private notes for practice:

Three bean plants received the same soil and water. Plant A was by the window, B was farther into the room, and C was in a cupboard. After a week A was tallest. We did not measure the actual light intensity.

Name the task and source

Start: “Using only the notes below, organise a short experiment summary.” This states the job and limits the evidence. Paste the notes after the instruction. Without them, the model may construct a different experiment.

Name the audience and shape

Add: “Write for a classmate who has not seen the experiment. Use four headings: Question, What we changed, What we observed, and What we cannot conclude.” Now you can check the response mechanically: it should contain four headings and explain the experiment for that audience.

State boundaries

Add: “Do not invent measurements. Keep uncertainty explicit. Do not say the experiment proves a universal rule.” These constraints matter because our notes contain no heights, no measured light levels, and only one plant per condition.

The complete prompt is:

Using only the notes below, organise a short experiment summary for a classmate who has not seen it. Use four headings: Question, What we changed, What we observed, and What we cannot conclude. Do not invent measurements. Keep uncertainty explicit. Do not say the experiment proves a universal rule.

[Paste the notes here.]

The bracketed instruction is a placeholder for your source, not something to leave in the finished request.

Inspect an illustrative answer

A reasonable “What we observed” section says that A was tallest after a week. A bad section might say “A grew 8 cm, proving sunlight doubles growth.” That adds both an invented measurement and an unsupported conclusion.

Useful feedback is specific:

Remove the 8 cm measurement because it is not in the notes. Replace the claim of proof with an explanation that this small experiment suggests a possible relationship but does not establish it reliably.

You are evaluating evidence and revising a particular defect. You are not just asking the model to sound more scientific.

Show an example when format is hard to describe

Suppose you want vocabulary entries in a fixed pattern. Show a neutral example:

Term: variable | Meaning: a name for a value | Example: score = 3

Then ask for new entries in the same pattern. This demonstrates separators and length more precisely than “make it neat”. Verify the meanings separately; imitating a layout is not a fact check.

Ask for useful explanations

Ask the system to identify assumptions, show a calculation you can inspect, or explain which source supports a conclusion. Do not assume a generated explanation is a faithful record of the model's internal process. Its purpose for you is to expose claims you can test.

A role such as “patient tutor” can guide tone. It does not grant qualifications or make an answer authoritative. A long prompt is not automatically a good prompt either; remove instructions that do not help the job.

Your reusable pattern

State the job, provide the relevant material, name the audience and shape, state important boundaries, and inspect the output. Then revise the part that missed the goal. You should be able to explain the purpose of each instruction instead of collecting magic phrases.

Assemble a request line by line

Start with a job: “Help me plan a revision session.” Add the learner: “I am a beginner learning fractions.” Add the time: “I have twenty minutes.” Add the form: “Give three activities with minutes beside each.” Add the boundary: “Use examples I can do with pencil and paper; do not solve my homework questions.”

Each line removes a different ambiguity. Audience changes vocabulary; duration changes scope; output shape makes the plan inspectable; constraints rule out unavailable resources. No line guarantees that the arithmetic or teaching advice is correct, so read the resulting plan critically.

Personalise the subject and time while keeping the other pieces coherent. If the activities total thirty minutes, identify the concrete mismatch in a follow-up: “These times total thirty minutes; revise them to total twenty.” That request says what failed and what a successful correction would look like.