Prompt patterns that work

Once the basics are in place, a handful of reusable patterns will take you most of the way. Each one solves a specific problem. Learn to recognize the problem, and the pattern follows.

Pattern 1: show an example

Problem: the model keeps missing the style or format you have in mind.

Fix: give one to three examples of the input and the output you want. This is called few-shot prompting. One example is often enough to fix tone and format.

Turn each support ticket into a one-line summary in this style.

Ticket: "Hi, I can't log in since the update, it says invalid token."
Summary: Login fails with "invalid token" after update.

Ticket: "The export button does nothing when I click it on Safari."
Summary: Export button unresponsive in Safari.

Ticket: "{paste new ticket here}"
Summary:

Make the examples varied, and make them correct. The model copies mistakes as faithfully as it copies style.

Pattern 2: separate instructions from material

Problem: the model confuses your instructions with the text you want it to work on, especially with long pasted content.

Fix: mark the boundaries clearly with labels, quote marks, triple dashes or simple tags.

Summarize the document below in five bullets for a manager.

<document>
...pasted text...
</document>

Consistent labels also help you reuse the same prompt, and they reduce the chance that instructions hidden inside pasted text are followed by mistake.

Pattern 3: ask for the steps

Problem: multi-step questions such as calculations, comparisons or planning give wrong answers when the model jumps to a conclusion.

Fix: ask it to work step by step before answering, or use a reasoning mode if your assistant has one. Writing the steps gives the model room to work, and gives you something to check. This idea is often called chain of thought.

Decide which of these three vendors best fits our needs. First list our
requirements, then score each vendor against each requirement, then give your
recommendation and the main reason.

On assistants with built-in reasoning, you often do not need to ask: the model already thinks first. For simple questions, steps just add length.

Pattern 4: specify the output format

Problem: the answer is correct but cannot be used as it is.

Fix: describe the exact shape: a table with named columns, a numbered list, JSON with given keys, a message under 60 words.

Extract the following from the email below and return only JSON:
{"customer": string, "issue": string, "urgency": "low" | "medium" | "high"}
If a field is missing, use null.

Structured output is what makes prompts usable inside scripts and tools, which you will do in stage 2.

Pattern 5: break big tasks into steps (chaining)

Problem: one huge prompt gives a mushy result.

Fix: split the work. First extract the facts, then analyze them, then write the final text. Each step is short, and you can check each one.

A small example: (1) "List the key claims in this report", (2) "For each claim, say what evidence the report gives", (3) "Write a one-page summary that separates well-supported claims from weak ones".

Pattern 6: let it ask you

Problem: you are not sure what information the model needs.

Fix: tell it to ask first.

I want to plan a migration from our on-premises file server to cloud storage.
Before you suggest anything, ask me up to five questions you need answered.

Pattern 7: draft, critique, revise

Problem: the first answer is acceptable but not good.

Fix: have the model review its own work against criteria you give, then rewrite.

Review your answer against these criteria: accurate, no jargon, under 150
words, includes one concrete example. List problems, then give a revised version.

Self-review catches style and completeness problems. It is much less reliable at catching factual errors, which is why you still check facts yourself (lesson 6).

Keep what works

When a prompt works well, save it. A short library of reusable prompts with blanks for the details is worth more than any trick. You will build one in the capstone.

Go deeper (optional)

AI prompt engineering: A deep diveAnthropic · 77 min

Resources

Check yourself

1. Pick one of the seven patterns and apply it to a real task. What changed in the answer?
2. For each problem (wrong format, wrong style, wrong conclusion, too big), name the pattern that fixes it.
3. Why is asking a model to critique its own answer weaker for facts than for style?