The five prompt patterns every professional needs

Most professional work that AI can help with fits one of five reusable patterns. Learn the five and you have a vocabulary for the next decade of AI work.

This chapter introduces each pattern with one worked example. The book's appendix has 50+ templates across HR, legal, finance, engineering, customer support, government, and healthcare.

Pattern 1: Persona + Task + Output

The foundational pattern from Chapter 1. You set a Role, state a Task, define Context, specify Output format.

When to use: any task where you want the model to produce a deliverable directly.

Example:

> You are a senior contracts attorney specializing in federal procurement. > > Task: Review the attached MSA draft for terms that should be revised before signing. Focus on liability cap, IP assignment, and indemnity. > > Output as a Markdown table with three columns: section reference, issue, recommended revision. > > MSA draft: > [paste]

Pattern 2: Few-Shot

Show the model 2-5 examples of the input-output mapping you want. The model patterns from the examples.

When to use: tasks where the format is fiddly or unconventional, or where you want a specific tone.

Example:

> Convert each customer email to a one-sentence "what they actually want" summary. > > Email 1: > "Hi! I bought your widget last week and it's not working. The button on top doesn't depress. I've tried turning it off and on. Help!" > Summary 1: Button-on-top is stuck on widget purchased ~1 week ago. > > Email 2: > "Your support page said to contact you. I want a refund for order #4521. I've moved abroad and the warranty won't apply." > Summary 2: Refund request for order 4521 due to international move. > > Email 3: > "[the new email to classify]" > Summary 3:

The model will follow the pattern. You can chain 50 emails through this and get 50 clean summaries.

Pattern 3: Chain-of-Thought

Ask the model to think step by step BEFORE giving its answer. The model produces better answers when it can show its work.

When to use: reasoning tasks, math, multi-step decisions, anything where the model might shortcut.

Example:

> A federal contracting officer needs to determine if a proposed subcontract triggers FAR 19.5 limitations. > > Subcontract value: $2.3M > Prime contract value: $4.8M > Subcontract scope: cybersecurity awareness training > Prime contract scope: cybersecurity assessment and training services > Set-aside type: Total Small Business > > Think step by step. First identify the applicable FAR 19.5 percentage rule. Then calculate the percentage of work being subcontracted. Then determine whether the subcontract complies. Only then give your final answer.

The "Think step by step" phrase is load-bearing. Without it, the model may guess. With it, the model usually reasons correctly.

Pattern 4: Output-Format Spec

The full-power version of the "Output format" hint from Pattern 1. You give the model a JSON schema or structured-output spec.

When to use: anything you want to pipe into other software (spreadsheet, database, automation).

Example:

> Extract the following fields from the attached SOW into the JSON schema below. > > Schema: > { > "period_of_performance": { > "start_date": "YYYY-MM-DD", > "end_date": "YYYY-MM-DD", > "option_years": number > }, > "ceiling_value_usd": number, > "place_of_performance": [string], > "key_personnel": [{"name": string, "role": string}], > "deliverables": [{"name": string, "due_date": "YYYY-MM-DD"}] > } > > SOW: > [paste] > > Output the JSON object only. No surrounding prose.

You will get parseable JSON every time. (Use a model that supports JSON-mode for higher reliability.)

Pattern 5: Context Embedding

The reverse of Patterns 1-4. Instead of asking the model to produce a deliverable, you ask it to ANALYZE or QUESTION the context.

When to use: when the context is large and you want the model as a reading assistant rather than a producer.

Example:

> Below is a 40-page solicitation for an AI training procurement at the Department of Energy. > > Read carefully. Then answer these specific questions: > > 1. What's the cited NAICS code? > 2. What's the set-aside type, if any? > 3. What's the questions-deadline date? > 4. What's the response-deadline date? > 5. Are there any unusual evaluation factors (beyond price + technical)? > 6. Are there any Section L requirements that would disqualify a small business? > > Solicitation: > [paste] > > Output as numbered list.

The model becomes a reading assistant for documents that would otherwise take a human 45 minutes to skim.

Which pattern to pick

| You want… | Use | |---|---| | A deliverable in a specific format | Pattern 1 (Persona + Task + Output) | | Consistent format across many inputs | Pattern 2 (Few-Shot) | | Better reasoning on hard problems | Pattern 3 (Chain-of-Thought) | | Structured data for automation | Pattern 4 (Output-Format Spec) | | Quick analysis of a big document | Pattern 5 (Context Embedding) |

These five compose. A prompt can be Persona + Few-Shot + Output Spec. Or Chain-of-Thought + Output Spec. The patterns are building blocks.

The rest of this chapter has 5 detailed worked examples — one per pattern, drawn from real customer engagements — including the failure modes that led to each pattern's discovery.

---

*This is an excerpt from Chapter 5 of Prompt to Product. The full chapter includes 5 worked examples, 12 anti-patterns to avoid, and the appendix referenced above includes 50+ templates across professional domains.*

*Buy direct ($19.95) · Amazon*