Clear instructions
The five-part anatomy of a working prompt: role, task, facts, constraints, format — and why specificity beats magic words.
There are no magic words. Prompting well is just briefing well — the same skill as delegating to a smart new colleague who knows nothing about your company and takes everything literally. Good briefs share five parts — and you already used four of the five in Module 1's lab; this lesson just names the anatomy.
The five-part anatomy
- Role — who the model should be. 'You are an experienced HR manager reviewing a job description' loads the right vocabulary, priorities, and tone.
- Task — one clear verb. Summarize, draft, compare, critique, extract. If you can't name the verb, the model can't either.
- Facts — everything it needs to know that isn't public knowledge. Paste the email thread, the data, the notes. The model can't read your mind, your inbox, or your intranet.
- Constraints — length, tone, audience, what to avoid. 'Under 200 words, no jargon, don't mention the vendor by name.'
- Format — what the output should look like. Bullets? A table? An email with subject line? Numbered steps? Say so, and you'll never reformat by hand again.
Here's Sam applying all five to a real Harbor Lane task:
You are an operations analyst at a mid-size retailer. [ROLE] Draft the weekly ops summary for store managers. [TASK] This week's facts: warehouse B cleared its backlog (was 3 days, now same-day). Two delivery trucks were out of service Tue–Wed, causing 41 late deliveries in the Northeast. The new returns portal launches Monday. [FACTS] Tone: direct, no corporate filler. Under 150 words. Don't assign blame for the truck issue. [CONSTRAINTS] Format: three sections titled Wins, Issues, Coming up — each with 1–3 bullets. [FORMAT]
The bracketed labels are for you, not the model — remove them in real use. Run it, then delete the FACTS section and run again. Watch the model invent a plausible-but-fake week. Facts are the load-bearing part.
The specificity dial
Vague prompts don't produce errors — they produce averages. 'Make it better' returns the average of everything 'better' means on the internet. 'Make it shorter, cut the second paragraph, and end with a clear ask for budget approval by Friday' returns your edit. When output disappoints, the fix is almost never a cleverer phrase; it's a more specific brief.
Before sending a prompt, ask: if I sent exactly this to a capable temp on their first day, would they produce what I want? If they'd have to guess anything important, the model will guess it too — just faster and more confidently.