A prompt pattern is a reusable, named solution to a recurring problem in prompting a language model, documented with its intent, its structure and the consequences of applying it. In the field of artificial intelligence, the idea borrows directly from software design patterns: instead of describing one good prompt, a pattern describes a family of prompts that solve the same kind of problem across tasks, domains and models.
Collections of such descriptions are known as a prompt pattern catalog. Practitioners use them to share proven techniques, to give teams a common vocabulary for prompt design and to make prompt behavior more predictable when models, data or requirements change.
How a prompt pattern is structured
A prompt pattern is structured as a short specification rather than a finished prompt. It names the problem, explains why naive prompts fail at it and lists the fundamental contextual statements that any concrete prompt must contain for the pattern to work. The concrete wording is left to the author, because the same intent can be expressed in many phrasings.
Most catalogs describe each pattern with a fixed set of fields. The table below shows the fields that appear in nearly every well-documented pattern and the question each one answers.
| Field | Question it answers | Example content |
|---|---|---|
| Name | How do practitioners refer to it | Persona, Flipped interaction |
| Intent and context | Which problem does it solve | Model answers lack domain perspective |
| Motivation | Why do ordinary prompts fail here | Generic answers ignore the audience |
| Structure | Which statements must the prompt include | Act as X; respond as X would |
| Example implementation | What does a concrete prompt look like | A sample instruction and output |
| Consequences | What are the benefits and side effects | Better focus, risk of invented detail |
The consequences field is the part most often skipped in informal collections and the part experienced engineers value most. It records known failure modes, interactions with other patterns and the conditions under which the pattern stops helping, which turns a tip into engineering knowledge.
Main families of prompt patterns
Prompt patterns are commonly grouped by the aspect of the interaction they control: what goes into the model, how it should reason, what comes out and how the conversation evolves. Grouping helps authors find candidate patterns quickly and spot combinations that work well together.
| Family | Representative patterns | Primary goal |
|---|---|---|
| Input semantics | Meta language creation, glossary definition | Teach the model new notation |
| Output customization | Persona, template, output format instruction | Shape tone, structure and format |
| Error identification | Fact check list, reflection | Expose claims that need verification |
| Prompt improvement | Question refinement, alternative approaches | Let the model improve the request |
| Interaction | Flipped interaction, game play | Change who drives the dialogue |
| Context control | Context manager, step-back prompting | Decide what the model considers |
| Reasoning | Chain of thought, self-consistency, decomposition | Improve multi-step accuracy |
Several neighboring entries in prompting, such as role prompting, least-to-most prompting or chain-of-verification, can be read as individual patterns inside these families. The pattern view adds value by describing when each one applies and how it combines with others, rather than treating each technique as an isolated trick.
Prompt pattern compared with related concepts
A prompt pattern differs from neighboring concepts mainly in its level of abstraction. It sits above a concrete prompt and a fill-in template, but below a broad framework that organizes an entire prompt or workflow. Confusing these levels leads to catalogs that mix reusable ideas with one-off prompts.
| Concept | Level of abstraction | What it gives the author |
|---|---|---|
| Prompt pattern | Problem and solution idea | Reusable reasoning about design choices |
| Prompt template | Parameterized text | Fixed wording with variable slots |
| Prompt framework | Checklist for a whole prompt | Sections such as context, objective, style |
| Prompt library | Stored collection of prompts | Ready-made prompts for known tasks |
| Prompting technique | Method, often from research | A specific procedure with measured gains |
In practice the levels connect: a team selects one or more patterns, implements them as a prompt template and stores that template in a prompt library under version control. The CO-STAR framework, for instance, organizes a whole prompt, while a persona pattern describes only one design decision within it.
Applying and combining prompt patterns
Applying a prompt pattern means mapping its required statements onto the concrete task and then testing the result. Patterns are rarely used alone; production prompts typically combine two to five patterns, such as a persona, a template for the output and a reflection step that asks the model to list its assumptions.
The decision table below connects frequent problems with a suitable starting pattern. It is a starting point for experimentation, not a guarantee of improvement, because effects vary by model family and task.
| Problem observed | Starting pattern | Reason |
|---|---|---|
| Output breaks a parser | Template or output format instruction | Fixes field order and syntax |
| Answers are too generic | Persona or audience persona | Adds perspective and register |
| User cannot specify needs | Flipped interaction | Model asks clarifying questions first |
| Multi-step errors | Chain of thought or decomposition | Exposes intermediate steps |
| Unverified factual claims | Fact check list | Lists claims for human review |
| Vague original request | Question refinement | Model proposes a sharper question |
- Combine patterns in a deliberate order: context first, then role, task, constraints and output format.
- Check for conflicting instructions, for example a terse template combined with a request for detailed reasoning.
- Evaluate each change on a fixed test set instead of judging by a single impressive answer.
- Record the pattern names in prompt comments so reviewers understand why the wording exists.
Limitations and common mistakes
The main limitation of prompt patterns is that their effects are model-dependent. A pattern that improves accuracy on one model can be neutral or harmful on another, and newer models often perform reasoning internally, which reduces the benefit of explicit reasoning patterns. Catalogs rarely include quantitative evidence, so their claims should be treated as hypotheses to verify.
A second issue is prompt sensitivity: small wording changes inside a pattern can shift results noticeably. Treating a pattern as a magic phrase, rather than as a design idea that still needs evaluation, is the most common error among new practitioners.
| Mistake | Consequence | Better practice |
|---|---|---|
| Stacking too many patterns | Long, contradictory prompts | Keep only patterns with measured value |
| Copying wording verbatim | Poor fit to the task | Adapt the required statements |
| Skipping evaluation | Hidden regressions | Use a regression test set |
| Ignoring model updates | Prompt drift over time | Retest after each model change |
| Trusting persona expertise | Confident but invented detail | Pair with verification patterns |
The short version
A prompt pattern is a proven recipe idea for asking an AI model something. It does not give you the exact words. It tells you which ingredients your request needs to solve a certain kind of problem, such as getting a neat table or getting the model to ask you questions first.
Think of it like the advice an experienced cook passes on: always rest the meat, always taste before serving. You still choose the dish. The pattern gives you the reliable move, and you adapt it to your own situation and then check whether it actually worked.
Where it makes a difference
A support team at a mid-sized software company uses a language model to draft replies to customer tickets. The first version, a single long instruction, produced friendly but inconsistent answers, and about one in five drafts missed the required ticket summary block that the help desk system parses automatically.
The team rebuilt the prompt from three documented patterns: a persona describing a senior support engineer, a template defining the exact reply structure with a summary block, and a fact check list asking the model to list every product claim it made. They tested the new version on 200 archived tickets before rollout.
Parsing failures dropped to almost zero, and reviewers could check the listed claims in seconds instead of reading every draft line by line. Because the pattern names were recorded in the prompt file, a new engineer later understood why each section existed and could swap the model without rewriting the design.