Prompting like a professional

Prompting isn't magic words. It's a set of patterns that make the model do reliable work. Master five patterns and you'll get better output than 95% of AI users.

The anatomy of a great prompt

Every strong prompt has four parts. Missing any of them is where things go wrong.

1. Role. Who is the model right now?
2. Context. What material is it working with?
3. Task. What exactly should it do?
4. Format. What shape should the answer take?

A minimal example

Prompt# Role
You are a senior proposal writer for a design agency that specializes
in accessible government websites.

# Context
Below is an excerpt from a California city RFP for a new tourism website.
<RFP>
{paste RFP text here}
</RFP>

# Task
Extract every stated requirement. Ignore boilerplate. Distinguish "must"
from "should" or "may."

# Format
Return a Markdown table with columns: Requirement, Priority (Must/Should/May),
Section reference.

Five patterns worth memorizing

Few-shot in practice

PromptClassify each inbound email as: NEW_LEAD, EXISTING_CLIENT, RECRUITMENT, SPAM.

# Examples
Email: "Hi, saw your work for the City of Fremont, we're looking for a redesign."
Class: NEW_LEAD

Email: "Hi — invoice #4293 for August still open, can you resend?"
Class: EXISTING_CLIENT

Email: "Growth Marketing Intern here, want to help scale your firm."
Class: SPAM

# Now classify:
Email: "{email_body}"
Class:

Structured output

When the answer needs to be consumed by code, force JSON:

PromptExtract the client, budget range, and deadline from the message below.

Return ONLY a JSON object matching this schema:
{
  "client_name": string,
  "budget_usd": { "min": number|null, "max": number|null },
  "deadline_iso": string|null,   // ISO 8601 date, or null if unspecified
  "confidence": "high" | "medium" | "low"
}

If a field is unclear, use null. Do not invent values.

Message:
<email>
{email_body}
</email>

Both OpenAI and Anthropic support "structured outputs" natively — you can pass a JSON schema and the model is guaranteed to return valid JSON. Use it whenever a downstream script consumes the result.

Anti-patterns to unlearn

• "Be creative" — meaningless. Give constraints instead: "3 headlines, each under 8 words, no puns."
• "Please make it better" — better how? Specify: "shorter, more concrete, remove passive voice."
• Dumping 20 tasks at once — split into steps. The model will half-do all of them.
• Trusting one shot — for anything important, sample twice, compare, or use evaluator-optimizer (Module 4).