8 Prompt Engineering Patterns: For Research, Work, and Personal Use

A prompt is really a spec

Most people write to ChatGPT the way they would write to a friend. They say "write me an email," the response comes back short and generic. Then they start correcting it with "this is too generic." That loop eats hours, and the output still fails to satisfy.

The right approach: a prompt is thought of like a software spec. Which task, in which format the output, in which tone, within which limits, with which references. The clearer the spec, the more usable the returned output.

In this article I walk through the eight patterns. We will look at when each one produces the most value and at an example prompt for each. The patterns do not block each other; as they stack on top of one another, quality compounds.

Pattern 1: Role Assignment

We tell the model what role it is in. "You are X, help me with Y."

Why it works: The model uses the role information as a perspective filter. To the same question, "you are a lawyer" and "you are a journalist" produce very different answers.

Example prompt:

You are a UX researcher with 15 years of experience. For a beginner,
write a practical 5 step roadmap on "how to plan a user interview."
Use hands-on language, not academic language.

Common mistake: Too general a role ("be an expert"). A specific role is better: "a CRO expert with 20 years of experience, focused on B2B SaaS."

Pattern 2: Output Structuring

Clearly defining the format of the output. Table, headed list, JSON, markdown, code, a specific template.

Why it works: In free format the model writes loosely. When the format is enforced, discipline increases, and it also becomes easier to parse.

Example prompt:

Evaluate the 5 ideas below. Give the output as this Markdown table:
| Idea | Market size (S/M/L) | Technical difficulty (1-5) | Competition (Low/Medium/High) | My recommendation (Go/Wait/Drop) | Reason (1 sentence) |

Ideas:
1. ...
2. ...

Advanced: Asking for JSON output (to process it programmatically later):

Respond in JSON format, with this schema:
{
  "summary": "string",
  "key_points": ["string"],
  "risk_score": 1-10,
  "next_steps": [{"action": "string", "priority": "high|medium|low"}]
}

Pattern 3: Few-shot Examples

Showing the model what you want by giving examples. One example, three examples, five examples.

Why it works: Models learn the pattern from examples very quickly. Giving a concrete example instead of an abstract description noticeably increases quality and consistency.

Example prompt:

I give you a product description. You produce a 30 word tweet-friendly
summary. In the style of the example below:

Example 1
Input: "These headphones, with active noise cancellation..."
Output: "For anyone who wants to sleep on the train: ANC works, battery
30 hours, mid-range price. Great unless you are an audio nerd."

Example 2
Input: "..."
Output: "..."

Now write for this:
Input: [new product description]

Its strength: Style, tone, terminology, length, and format are all conveyed through the example. While an abstract instruction ("make it fun but informative") falls short, three examples make it clear.

Pattern 4: Chain-of-Thought

Asking the model to think before it answers. "Think step by step," "first write the plan, then give the answer."

Why it works: On tasks that require complex reasoning (math, multi step analysis, chains of logic) accuracy increases by 20 to 40 percent. Modern models (GPT-4, Claude 3.5+, Gemini 1.5+) do this pattern internally as well, but asking for it explicitly still raises quality.

Example prompt:

Analyze this case study. First think through these steps in order:
1. Which numerical values can you extract from the given information?
2. Which assumptions do you need to make? State them explicitly.
3. Which calculation formula is appropriate?
4. Do the calculation step by step.
5. Interpret the result from a clinical/business perspective.

Then write your final answer clearly.

Case:
[case description]

Important: The model's "thinking" part should be visible in the output. If it is not visible, it may be a case where the model actually skipped a step, and you lose control.

Pattern 5: Decomposition

Breaking a large task into sub parts and requesting them in sequence.

Why it works: When too much is asked in a single prompt, models leave some of it incomplete. A decomposed prompt gives full attention to each part.

Example workflow:

Instead of a single prompt:

"Write me a blog post, on topic X, SEO friendly, 1500 words, with a catchy title, H2 subheadings, and a CTA at the end..."

Three separate prompts:

1. "Write 5 title suggestions on topic X for target audience Y, explaining why each one will work."
2. "For this title [chosen] produce an outline with 7 H2 subheadings. Each subheading with a 1 sentence summary."
3. "Write the article according to the outline. 200-300 words under each H2. A 1 paragraph CTA at the end."

At each step you keep control in your hands. If you did not like a title in the first step, you revise it, and the later steps do not go down the wrong path.

Pattern 6: Persona Refinement

Defining not just the model's role but also the depth of its perspective.

Why it works: "You are a designer" and "you are a visual designer, 25 years old, working freelance in Berlin, focused on minimalist design, who designs coffee shops" give very different output. The second, with its richness of character, produces more creative and specific content.

Example prompt:

You are a Gen Z consumer, 22 years old, studying sociology at university,
oversaturated with social media, and now looking for deeper content. You
have been told to listen to the podcast below. You are open minded but if
you get bored you say so honestly. What do you think after 5 minutes?

Podcast description: [content]

Use cases:

• Marketing message testing
• Content validation with a user persona
• Producing a counter perspective in academic debate
• Scenario/dialogue writing

Pattern 7: Constraint Setting

Drawing the limits clearly. A do-not list, word count, language level, numerical limits.

Why it works: By default the model writes verbose and general. Under the pressure of constraints it gives less but more concise output.

Example prompt:

Summarize this:
- Maximum 60 words
- Do not use technical jargon (for someone who is a university graduate but
  not an expert in this field)
- Do not use passive voice
- Do not use empty filler like "the important thing is," "notably,"
  "interestingly"
- If there is a numerical value, always state it

Advanced combined constraint:

Hard rules:
- Answer only in JSON format
- No more than 5 keys
- Each value a string or array

Style rules:
- Write in Turkish
- Do not use em-dashes, prefer commas or periods
- If there is a foreign word, give its Turkish equivalent in parentheses

Pattern 8: Iterative Refinement

Not accepting the first output, improving it with feedback.

Why it works: The first output is usually a 60 percent draft. Improving it brings it to 90 percent over 2 to 3 rounds. But refinement is also a prompting skill.

Example refinement round:

1. First prompt: "Write me an announcement email for a product launch."
2. First output: Generic, typical AI language.
3. Refine 1: "I liked these parts: [X]. Change these: make the tone more casual, speak from the you perspective, not the me perspective. Cut this sentence: 'The important thing is that...' This is a typical AI opening."
4. Second output: Better but still long.
5. Refine 2: "You caught the tone. Now shorten it by 30 percent. The opening sentence should hook the reader, the second sentence should deliver the value proposition directly."
6. Third output: Usable.

Advanced self-critique:

Now critique the text you wrote with the eye of a strict editor.
Which 3 weaknesses do you see? How would you fix each one?
Then write the corrected version.

This pattern is especially valuable in tasks that do not come together in one pass, like creative writing, manuscript drafts, and email.

Combining the patterns

A single pattern is nice, combined patterns are superior. In a typical high quality prompt, four or five patterns are present together:

[ROLE] You are a technical content editor with 10 years of experience,
focused on B2B SaaS.

[CONSTRAINT] You will revise the article below:
- Maximum 800 words (currently 1400)
- Do not use passive voice
- Each paragraph at most 3 sentences
- No em-dashes

[CHAIN-OF-THOUGHT] First read the article and answer these questions:
1. What is the main argument?
2. Which paragraphs are unnecessary?
3. Which sentences are repetition?
4. Which are the 3 weakest sentences?

[OUTPUT STRUCTURING] Then return in this structure:
## Analysis
[answers to the questions]

## Revised text
[800 word version]

## Editor's note
[3 sentences: what you did, what you sacrificed]

[FEW-SHOT] Reference for style: the paragraph below is the tone I like.
[example paragraph]

Now begin.

A prompt like this, with 5 patterns, raises the single-pass usable output rate from 30 percent to 80 percent.

Practical suggestions

Keep a prompt library. For tasks you use often, write 5 to 10 good prompts and save them in Notion or a text file. Do not rewrite, copy and paste.

Watch the difference between models. Give the same prompt to Claude, GPT-4, and Gemini, and see the difference. Every model has a "tone." Claude is strong in creative work, GPT-4 in code and chains of logic, Gemini in long document analysis.

Do not fall into the trap of having the AI write the prompt itself. When you ask "write me a good prompt," the AI produces a long, generic template. The real prompting skill is your ability to catch and add the details specific to your problem.

Learn its limits. Models hallucinate. They produce nonexistent references, wrong dates, wrong numbers. Treat every output as a draft that needs to be verified, not as fact.

Disciplined use of the eight patterns increases the value you get from AI by 5 to 10 times. A one day learning process brings years of savings.

For one on one support in prompt engineering, you can take a look at the individual mentorship option.