How-To: Write a Good System Prompt for a Specific Use Case

Step 1. — The structure of a good prompt

Reliable system prompts share five parts. Not every prompt needs all five, but most do:

1. •Role — who the AI is ("You are a meeting notes assistant").
2. •Context — the background it needs ("These notes go to a busy CEO who only reads bullets").
3. •Rules — the constraints, positive and negative ("Keep under 150 words. Don't use emoji.")
4. •Output format — exactly what shape the answer should take ("Return a bullet list with the subject line bolded").
5. •Examples — one or more samples of a good response, so the AI has a pattern to copy.

Step 2. — Email drafting prompt example

Here's a complete system prompt for an email drafting assistant:

You are an email drafting assistant. Write concise, professional emails. Always include a clear subject line prefixed with "Subject:". Keep each email under 150 words. Use a warm but professional tone. No exclamation marks. Sign off with "Best,". Here is an example of good output: Subject: Tuesday's design review — agenda attached Hi Sam, Quick heads-up that Tuesday's design review will cover the new checkout flow and the onboarding refresh. I've attached the agenda; please add anything you'd like discussed by end of day Monday. Best, [Your name]

Step 3. — Meeting notes prompt example

You are a meeting notes assistant. Convert the provided transcript into structured notes with these sections, in this order: - Attendees - Key Decisions - Action Items (each with an owner, in the format "Owner: name — task") - Next Steps Use bullet points. If the transcript doesn't mention an owner, write "Owner: unassigned". Do not include filler ("The meeting began at..."). Only list action items that were explicitly agreed, not opinions.

Step 4. — Test and iterate

A prompt is a hypothesis. Run it on 2-3 real inputs, judge the output, then adjust. A common loop:

1. •Run the prompt on a sample input.
2. •Score the output against what you actually wanted.
3. •Identify the gap (too long? wrong tone? missing section?).
4. •Add one instruction at a time, then re-run.

Example fix: a meeting-notes prompt that produced rambling paragraphs got a single line added — "Every bullet must be one sentence or less" — and the output tightened up immediately. Small, surgical edits usually beat full rewrites.

Step 5. — Common mistakes

• •Too vague: "Write good notes" gives the AI nothing to anchor on.
• •No format specified: the AI picks a format at random and it changes every time.
• •No examples: without a sample, "concise" or "professional" means whatever the model last saw.
• • Contradictory instructions: "Be detailed but brief" or "Be friendly but never informal" confuse the model. Pick one and add a concrete definition instead ("Brief = under 150 words").

Step 6. — Advanced techniques

• • Few-shot prompting: include 2-5 input/output examples in the prompt. More examples of good output measurably improve consistency.
• • Chain-of-thought: for reasoning tasks, add "Think step by step before answering" or "List your reasoning, then give the final answer." Useful for maths, logic, and multi-step analysis.
• • Output constraints: if you're feeding the result into another tool, specify the exact format — e.g. "Return valid JSON with keys subject, body, and tone. No prose outside the JSON."