Advanced Prompt Patterns
Beyond basics — 8 proven prompt patterns that consistently produce better output
Contents
Why patterns matter more than magic words
"Add more detail" is not a pattern. "Be concise" is not a pattern. These are vague instructions that work sometimes and fail others.
A prompt pattern is a reusable structure that reliably produces a specific type of output — because it changes how the model approaches the task, not just what it produces.
The 8 patterns below are not tricks. They work because they match how language models actually process information — by priming context, constraining output space, or enabling step-by-step reasoning.
Pattern 1 — 3 — The core patterns
1. Persona pattern: "You are a [specific expert] with [specific experience]. You write for [specific audience]." Sets the entire frame of reference before any task.
2. Output constraint pattern: "Write exactly [N] bullet points / words / examples. Not more, not fewer." Hard constraints prevent padding and force prioritisation.
3. Negative example pattern: Show what you do NOT want alongside what you do. "Here is a bad example: [X]. Here is a good example: [Y]. Now write [Z] following the good example's style."
These three alone improve most outputs by 40-60%.
Pattern 4 — 6 — The thinking patterns
4. Step-back pattern: Before answering, ask the model to identify the underlying principles. "Before answering, state the core principle or framework that applies to this question." Gets more systematic answers.
5. Devil's advocate pattern: "Give me the strongest possible argument against [your recommendation]." Use after you have a good answer to stress-test it.
6. Rubber duck pattern: "Explain your reasoning step by step as you work through this. Show your assumptions and flag anything you are uncertain about." Forces transparency in reasoning — errors become visible.
Pattern 7 — 8 — The output patterns
7. Template fill pattern: Provide the exact output structure you want with placeholders. The model fills in the template instead of choosing its own format. Eliminates formatting iteration entirely.
8. Layered refinement pattern: Generate → critique → revise, all in one prompt: "Write [X]. Then review your own output and list 3 specific improvements. Then rewrite it incorporating all 3 improvements. Only show me the final version."
This consistently outperforms a single-shot generation because the model is able to evaluate its own output — it just needs to be instructed to do so.