How to write a good prompt (a primer for non-engineers)

Prompts are the new spreadsheet formula. Most people write them once and accept whatever they get. A few people iterate, and their AI output is dramatically better. Below are the six principles + concrete before/after examples.

The six principles

1. Specify the task

Vague: "Help me with this email."

Specific: "Rewrite this email to be 30% shorter while keeping the apology, the next step, and the deadline."

The model can only do what you describe. Describe more.

2. Specify the audience

Vague: "Explain RAG."

Specific: "Explain RAG to a non-technical product manager who has heard the term but does not know how it differs from fine-tuning. They will use this in a meeting with their CTO."

Audience drives tone, depth, and vocabulary.

3. Specify the format

Vague: "Summarise this meeting."

Specific: "Summarise this meeting in 4 sections — Decisions, Action items (with owners), Open questions, Next meeting agenda. Use bullet points. Keep it under 150 words total."

Without format guidance, the AI defaults to "thoughtful essay." Often you want a list.

4. Specify constraints

Vague: "Write a job description for a senior engineer."

Specific: "Write a job description for a senior backend engineer at a 12-person AI startup. Must include: gender-neutral language, no specific years of experience required, 5 specific responsibilities, 3 'nice to haves,' a sentence on equity. Avoid: 'rockstar,' 'ninja,' 'fast-paced environment.'"

Constraints are how you avoid the default-ChatGPT-voice.

5. Show examples (highest-leverage move)

Telling the AI "be concise" works partially. Showing the AI 2-3 examples of "concise" works dramatically better.

Pattern: "Here are three good examples of what I want: [example 1], [example 2], [example 3]. Now produce the same for [new input]."

This is called few-shot prompting. It is the biggest leverage point in prompt engineering. Use it everywhere you care about quality.

6. Iterate

The first prompt is a hypothesis. Run it on 10 cases. Look at the output. Identify what is consistently wrong. Adjust. Re-test.

Two iterations usually get you to "good enough." Five iterations get you to "excellent for our use case." More than ten suggests you are at the wrong abstraction — maybe you need to break the task into smaller steps, or your model choice is wrong.

Concrete before/after

Bad prompt

"Write a marketing email."

Better prompt

"Write a 90-word marketing email announcing our new product launch. Audience: existing customers, mostly small business owners in India. Tone: warm and direct, like a personal note from the founder. Include: 1 specific benefit, 1 specific feature, a clear CTA to book a 15-minute call. Avoid: jargon, exclamation marks, the word 'excited.' Format: subject line + body, no salutation. Here is an email we sent last quarter that worked: [paste example]."

The second prompt has the task, audience, format, constraints, and an example. The output quality difference is order-of-magnitude.

Patterns that work in production

Role + task + format

"You are a senior tax accountant in India. Given the following invoice data, produce a draft GSTR-1 row in the official format."

Chain of thought

"Before giving the answer, explain your reasoning step by step. Then provide the final answer in the format above."

Especially useful for math, logic, multi-step tasks. The model thinks better when it shows its work.

Refusal anchors

"If the input does not contain enough information to answer with confidence, respond with 'INSUFFICIENT DATA' followed by what you would need."

This single sentence stops a huge category of hallucinations.

Citation requirement

"For every claim in your answer, cite the source document and the paragraph number. If you cannot cite, do not make the claim."

Forces grounded responses. Essential when AI is reading documents.

Things people get wrong

"Just add 'be more accurate' to the prompt"

Adjective-stacking does not work. "Be more accurate, helpful, concise, clear" is noise. The model already wants to be all those things. Be specific about what good looks like, with examples.

Mega-prompts

Long prompts that try to cover every edge case usually degrade the model's behaviour. Better: a short, focused prompt + structured task decomposition.

Skipping evaluation

Writing a prompt without testing it on representative cases is gambling. Always test on a small eval set before shipping.

For non-engineers using AI day-to-day

Three habits that pay off:

1. Keep a prompt library. Every time a prompt works well, save it. Reuse for similar tasks.
2. Include a "good example" from your past work in every prompt where quality matters. Past you is the best teacher.
3. End with a question. "Does this answer match what I asked for? If not, what would you change?" The model often catches its own gaps.

What this means for you

• Specific beats clever. The boring 5-part brief wins.
• Show 2-3 examples. This is the highest-leverage move.
• Iterate on a real eval set. The first prompt is never the best.
• Refusal anchors and citation requirements stop most hallucinations.
• Read our 2026 foundation models guide — the right prompt for Claude differs from the right prompt for GPT.

Want help systematising prompts across your team? Book a 30-minute call and we will show you how we do it.