Prompting basics: how to get a useful answer the first time

Prompting is not a trick. It is writing a clear brief for a very fast, very literal colleague who has no memory of your last conversation. The colleague will do exactly what you ask, so the quality of the answer is mostly the quality of the ask.

The five parts

• Role — who the model should be. "You are an experienced HR manager" changes vocabulary and judgement.
• Task — the verb. Summarise, rewrite, compare, list, draft, critique. One task per prompt when possible.
• Context — the material and the situation. Paste the document. Say who the audience is.
• Constraints — length, tone, what to avoid, what must be included.
• Format — bullets, table, JSON, a subject line plus body, three options. Say it explicitly.

Worked example 1: the email

Before: "Write an email to my landlord about the broken heater." You get a generic, slightly too polite email with placeholders.

After: "You are a tenant writing to a landlord. Draft an email requesting repair of a heater that has not worked since Monday. Context: this is the second request; the first was ignored. Constraints: firm but polite, under 120 words, mention that the lease requires repairs within 14 days. Format: subject line, then body." You get an email you can send with one edit.

Worked example 2: the summary

Before: "Summarise this." followed by a 20-page PDF. You get a summary of whatever the model decided mattered, at whatever length it chose.

After: "Summarise the attached report for a finance director who has five minutes. Format: three bullets on findings, two on risks, one on the recommended decision. Quote the exact figures where they appear; do not estimate." The last sentence is the important one: it tells the model to prefer the source over its own guess.

Worked example 3: the code fix

Before: "My Python script does not work." The model asks what the error is, and you have spent a turn.

After: "Here is a Python function and the full traceback. Explain the cause in two sentences, then give the corrected function only. Do not change the function signature." Paste both. One turn, one answer, and the constraint prevents the model from rewriting things you did not ask about.

Two habits for ChatGPT specifically

First, iterate in the same chat. "Shorter." "More formal." "Now as a table." Follow-ups are cheap and the model already has the context. Second, when you want the model to think harder, say so: "Consider three approaches before answering" or switch to a reasoning-capable model in the picker. Do not pay for a heavy model just to rewrite an email.