What is a prompt, really?

A prompt is a contract between you and the model.

That sentence sounds abstract. Stay with it. By the end of this chapter you'll see why it's the most useful frame.

A Google search is a guess. You type three or four words. Google guesses what you want. Sometimes it nails it. Sometimes it doesn't. The asymmetry of effort is wild — you spend two seconds typing, Google spends a billion CPU-cycles guessing.

A prompt is the opposite. A prompt asks YOU to be specific. The model is fast. The model is patient. The model will read your full paragraph of instructions. Most professionals waste this asymmetry by typing four-word Google queries into ChatGPT.

The four parts of any useful prompt

Every prompt that produces a good result has four parts. Most prompts that produce bad results are missing one or more.

1. Role

Tell the model what role to take. Not who they ARE — what role they should occupy for THIS task.

Bad: "You are an AI assistant." Good: "You are an experienced HR analyst at a mid-sized federal contractor reviewing this candidate's resume against the job description."

The first sentence gives the model no instruction. The second sentence anchors the model in a specific perspective with specific knowledge. The output quality difference is dramatic.

2. Task

State the task in a single declarative sentence.

Bad: "Help me with this candidate." Good: "Identify the three biggest gaps between this candidate's resume and the job description, and recommend whether to advance to a phone screen."

The second version tells the model exactly what to produce. Three gaps. A recommendation. Stop.

3. Context

Give the model the inputs. This is the part most professionals skip. They write a brilliant Role + Task and then say "you can find the resume below" and paste 200 words of resume. Wrong — paste the WHOLE resume. Models are good at filtering signal from noise. Give them more, not less.

Two rules:

• Paste the full thing. Not your summary of the thing. The full thing.
• Distinguish the inputs. Use markdown headings or bracketed labels. ["Resume:" ... "Job description:" ...]. The model parses this structurally.

4. Output format

Tell the model exactly how to format the answer. This is the secret weapon.

Bad: "Tell me what you think." Good: "Output as a JSON object with three fields: gaps (array of strings, max 3), recommendation (one of advance / decline / borderline), and reasoning (one paragraph)."

Now you can feed the output into a spreadsheet. You can compare across candidates. You can build a workflow on top of it. Without the format spec, you get a 4-paragraph essay that requires human reading every time.

Putting it together

Here is the same task written two ways. The first is what most people write. The second uses the four parts.

Version 1 (most people)

> Look at this resume and tell me if they're a good fit for the job.

You will get a paragraph of vague positive feedback. Useless.

Version 2 (the contract)

> You are an experienced HR analyst at a mid-sized federal contractor. > > Task: Identify the three biggest gaps between the candidate's resume and the job description below, and recommend advance / decline / borderline. > > Output as JSON: > { > gaps: [<string>, <string>, <string>], > recommendation: "advance" | "decline" | "borderline", > reasoning: "<one paragraph>" > } > > Resume: > [PASTE FULL RESUME] > > Job description: > [PASTE FULL JD]

Run Version 2 and you get a workable, comparable, sharable answer in 8 seconds. Run it across 50 candidates in a row and you have a screening pipeline.

This is the difference between using AI like a search engine and using AI like an employee.

Why "contract"?

A contract specifies:

• Who does what (Role + Task)
• What materials are exchanged (Context)
• What deliverable is produced (Output format)

That's exactly what a prompt is. A short, written, binding agreement between you and the model about what work to do.

The corollary: a prompt that succeeds reliably is one that someone OTHER THAN YOU can re-run and get a similar result. If your prompt depends on the context being in your head, it isn't a reusable contract. It's a one-off.

The Five Prompt Patterns in Chapter 5 are five contract templates that work for ~70% of typical professional tasks. Use them. Adapt them. Don't reinvent.

---

*This is an excerpt from Chapter 1 of Prompt to Product. The full chapter continues with worked examples for HR, legal, finance, and engineering use-cases, plus the 12 most common failure modes when one of the four parts is missing or weak.*

*The full book is available direct ($19.95) or on Amazon.*