The four-part prompt

Almost every prompt that ships useful work has the same four parts. The Skills in this Playbook all follow this shape. Once you see it, you'll notice it everywhere — and you'll notice which prompts are missing one of the parts and producing mush as a result.

Here's the honest version of the lesson, including the one part most online prompt guides oversell.

Role · Context · Task · Output format

Role. Tell the model who it's being. One sentence.

"You are a marketing analyst at a mid-market B2B SaaS company. Your audience is the VP of Marketing."

This anchors voice and vocabulary. It is the weakest of the four parts. Recent research has shown that role-prompting (the "you are an expert at X" line) barely moves accuracy on most knowledge work — it mostly shifts tone. Useful, not magical. Treat it as a style anchor, not a performance unlock.

Context. What does the model need to know to do the job? This is where most of the value lives.

"The attached file is October campaign data exported from HubSpot, paid-search channel only. Our fiscal year starts February. We define MoM as current-month vs prior-month, not vs same-month-last-year."

Names of systems, names of files, names of people's titles, what time window, what definitions, what segment. Anything a smart new hire would need to ask before they started — answer it up front.

Task. What you want done. Verb-first, specific.

"Identify the three campaigns with the steepest MoM drop in cost-per-MQL. Write a one-paragraph hypothesis per campaign about what's driving the decline. Flag any campaign that's also seen a >20% drop in landing-page conversion in the last 90 days."

Not "analyze this." Not "give me insights." Verb-first, with the noun the verb operates on. Specific output (one paragraph per campaign, three campaigns).

Output format. What the deliverable should look like.

"Word document. Three sections, one per distributor. Each section has the distributor name as a heading, a one-paragraph hypothesis, and a single 'recommended action' line. Conservative formatting — this goes to a CFO."

If you skip this, you'll get whatever the model felt like that day. Word count, headings, audience, tone, what to lead with, what to cut. Every detail you add here is a steering instruction the model uses.

Which of the four does the work

Part	Weight	What it controls
Role	Low	Tone, vocabulary
Context	High	Whether the model has the facts it needs
Task	High	Whether the model does the right thing
Output format	High	Whether the result is usable when it comes back

If you only have time to write one part well, write Context well. If you have time for two, Context + Output format. If three, add Task. Role is the part you write last and can sometimes skip entirely.

This is the opposite of what most online prompt guides emphasize — they spend most of their wordcount on the role line because it's the easiest part to write a tutorial about. Don't fall for it.

What this looks like in practice

A workable prompt structure when you're starting from scratch on something the Playbook doesn't have a skill for:

You are [role]. Your audience is [audience].

Inputs (attached to this conversation):
- [file name + one-line description]
- [file name + one-line description]

Definitions (the things this analysis depends on):
- [term]: [your shop's definition]
- [term]: [your shop's definition]

Task: [verb-first, specific, with the noun].

Output format: [filename]. [Structure: sections, headings, tables].
[Tone or audience constraint, if not obvious].

Edge cases:
- [what to do if a column is missing]
- [what to do if a number doesn't reconcile]

The five-paragraph template above will outperform most ad-hoc prompts on most knowledge work. It's the same skeleton every Skill page in this Playbook follows.

Common mistakes

• Front-loading the role, skimping on context. The model doesn't need a six-sentence introduction to who it is. It needs to know what data you've attached and how your business defines its terms.
• Verbs without nouns. "Summarize this" — summarize what about what audience in what length?
• Saying "format the output nicely." Nice is not a format. Tell it the file type, the section structure, and the audience.
• Treating the prompt as fire-and-forget. The first run will tell you which of the four parts you under-specified. Edit and re-run. Two minutes of editing beats five identical retries.

Where to go next

• Show, don't tell — the highest-leverage thing you can add to almost any prompt: a single example of the output shape you want.
• Make it a critic — what to do after the model has produced a draft.
• Decompose and verify — when one prompt isn't enough and you should split the work into stages.