The Anatomy of a Good Prompt
What you'll learn
- Name the five building blocks of a strong prompt
- Rewrite a vague prompt into a specific, structured one
- Understand why specifying output format saves the most time
The skeleton of a reliable prompt
Vague in, vague out. The fix is to give the model a clear structure. Five parts cover almost everything:
- Role — who the model should act as (“You are a careful technical writer”). Sets tone and default assumptions.
- Task — the single, explicit thing to do (“summarise the text below”).
- Context — the material to work from (the document, the data, the background).
- Format — exactly how the answer should look (“3 bullets, each under 15 words”).
- Constraints — the guardrails (“no jargon”, “don’t invent facts”, “British spelling”).
You won’t need all five every time, but naming them stops you from leaving out the ones that matter.
Why format is the secret weapon
The part people skip — format — is the one that saves the most time. If you don’t say “give me a table” or “return JSON with these keys,” you’ll get prose you then have to reshape. Specifying the shape up front means the output drops straight into your email, spreadsheet, or code. It also makes results consistent, which is what lets you reuse a prompt as a dependable tool rather than rolling the dice each time.
A quick before-and-after
- Before: “Can you help me summarise this report?”
- After: “You are a careful analyst. Summarise the report below in exactly 3 bullets, each under 15 words, no jargon. If a number isn’t in the text, don’t invent it. Report: ‘’‘…’‘’”
The second version tells the model its role, the task, the format, and the guardrails — and it will hand back something you can paste and send. The next lessons build on this: zero- and few-shot prompting, then system prompts and roles.
You are a careful technical writer.
Task: summarise the text below for a busy executive.
Format: exactly 3 bullet points, each under 15 words, no jargon.
If a fact is missing from the text, do not invent it.
Text: """{paste your text here}"""
• Revenue grew 12% last quarter, driven by enterprise renewals. • Support costs fell after the new self-service portal launched. • Main risk: one client is 30% of revenue and renewing in Q4.
Sample output — AI responses may vary.
Common mistakes
- Being polite but vague ('can you help with my essay?') instead of specific ('rewrite this paragraph in a formal tone, keep it under 80 words').
- Not specifying the output format — you then spend more time reformatting the answer than you saved.
- Dumping context without saying what to do with it. State the task explicitly, then provide the material.
How to check the AI's answer
- Read the output against your constraints: did it hit the word count, format, and 'don't invent facts' rules? If not, tighten the prompt, don't just fix the output by hand.
- Run the same structured prompt twice. A good prompt gives consistent structure each time; a vague one drifts.
Key terms
Test yourself
3 questions · answers reveal after you check.
1. Which element most reduces time spent reformatting answers?
Telling the model the format up front (bullets, JSON, word count) means you get usable output instead of reshaping it afterward.
2. A well-structured prompt usually includes…
Those five parts make intent explicit, which is what produces consistent, useful answers.
3. Why add 'if a fact is missing, do not invent it'?
Explicitly allowing 'not found' curbs the model's tendency to fill gaps with plausible-sounding inventions.