Prompt Patterns: Personas, Critique, Decomposition and Chaining

Experienced prompt writers rarely invent everything from scratch. They reuse patterns: proven ways of structuring a conversation with a model that work again and again across different topics. In this lesson you will learn six of them, and you will see how to combine them into a mini toolkit for your personal work assistant.

In the previous lesson you learned to control the shape of an answer. Now we look at the shape of the conversation: how to organize one or several prompts so the model does a better job than a single request could.

Pattern 1: the persona

You already know the "role" ingredient. A persona pushes it further: you describe who the model should be, including expertise, attitude and audience, to shift its vocabulary and priorities.

Act as an experienced editor who works with first-time authors.
You are encouraging but honest, and you explain every suggestion
in one sentence. Review the paragraph below.

Personas are helpful for tone and for perspective ("a skeptical investor," "a customer who has never used the product," "a patient teacher"). Two cautions:

• A persona changes the style and angle, but it does not give the model new knowledge. "Act as a doctor" does not make the answer medically reliable.
• Do not use a persona to make the model sound more certain than it is. Keep your reliability rules ("say when you are unsure") in place.

Pattern 2: critique, then improve

Models are often better at reviewing a text than at producing a perfect one in a single shot. This pattern splits the job into stages.

1. Ask for a first draft (or provide your own).
2. Ask for a critique against clear criteria.
3. Ask for a revised version that fixes the listed problems.

Here is my draft: [text]

Step 1: List the five biggest weaknesses of this draft in terms of
clarity, structure and tone for a skeptical reader.
Step 2: Rewrite the draft to fix those weaknesses.
Step 3: Briefly list what you changed.

The final line is valuable: it lets you see what changed and decide whether you agree. You can also ask for a critique only, and improve the text yourself.

Pattern 3: decomposition

Complex tasks fail when we ask for everything at once. Decomposition means breaking a large task into smaller steps, either inside one prompt or across several.

Example: "Plan and write a 10-minute talk about remote work" can become:

1. Define the audience and the one message to remember.
2. List five possible main points and select three.
3. Write an outline with timing.
4. Draft the opening and the closing.
5. Prepare speaker notes.

Each step is easier for the model and easier for you to check. You also keep control over the direction: you can change course after step 2 instead of discovering a wrong angle in a finished 2,000-word text.

Pattern 4: prompt chaining

Chaining takes decomposition across separate prompts: the output of one becomes the input of the next. Each link does one job well.

Prompt 1 (extract):  "List all dates, names and decisions in these notes."
Prompt 2 (organize): "Group those items by topic. Output a table."
Prompt 3 (write):    "Using the table, write a half-page recap
                      for the team, in a friendly tone."
Prompt 4 (check):    "Compare the recap with the table. List
                      anything in the recap that is not in the table."

Chaining has three big advantages: you can inspect and correct the intermediate results, you can reuse individual links in other chains, and the final prompt is short and focused. Its main cost is time and a bit more manual work. For important tasks, it is well worth it.

Pattern 5: clarifying questions

If a request is underspecified, a model will usually guess and move on. Ask it to question you first:

I want to write a short announcement for our volunteer event.
Before writing anything, ask me up to five questions you need
answered to do this well. Wait for my answers.

The model will ask about the date, the audience, the tone, the call to action and so on. This both improves the result and helps you realize what you had not decided yet. A variation is the "ask me first" rule in your custom instructions: "If key information is missing, ask me before answering."

Pattern 6: the checklist or criteria review

Before finalizing, ask the model to verify a result against criteria that you define:

Check this email against the following checklist and answer
yes or no for each item, with a short reason:
1. The request is stated in the first two sentences.
2. A deadline is mentioned.
3. The tone is polite and direct.
4. There is no jargon.
5. No personal data about third parties is included.

This turns quality control into a repeatable step. It also reminds you that the checklist is yours: the model verifies, but you decide what "good" means.

Combining patterns

These patterns are building blocks. A strong everyday workflow might look like this:

1. Clarifying questions to complete the brief.
2. Decomposition into an outline you approve.
3. Persona (a clear editor) to write the draft from your outline.
4. Critique, then improve to polish.
5. Checklist as a final review, then your own read-through.

When not to over-engineer

Patterns are tools, not rituals. If your task is simple ("Give me a synonym for 'important'"), a one-line prompt is fine. Use heavier patterns when the stakes or the complexity justify the extra steps. A good test: would a competent human colleague need a briefing, a draft and a review for this task? If so, so does the model.

Your turn: apply a pattern

Exercise 1. Choose a text you are writing (a message, a short report). Write a "critique, then improve" prompt with three criteria that matter for your reader, and run it.

Exercise 2 (project). Design a three-step chain for one of your project tasks, for example: (1) extract the key facts from some notes, (2) draft a message from those facts, (3) check the message against the facts. Write the three prompts, run them in order, and note where you intervened.

Prompt 1: "From the notes below, list only (a) what was completed,
(b) what is blocked, (c) what is next. Use bullet points and
do not add anything that is not in the notes."

Prompt 2: "Using these three lists, write a status update of at most
120 words for my manager. Tone: factual and calm."

Prompt 3: "Compare the update with the three lists. Point out any
statement in the update that is not supported by the lists."