Prompt types describe the shape of a prompt: what it contains and where. They are not exclusive. A real production prompt is usually a system prompt that assigns a role, gives instructions, includes a few examples, and requires a structured output. Knowing the types helps you decide which parts a task needs.
Quick comparison
| Type | What defines it | Use it when |
|---|---|---|
| Zero-shot | Instructions only, no examples | The task is common and easy to describe |
| One-shot | One example | The output format is the hard part |
| Few-shot | Two or more examples | The task depends on judgment or style |
| Role | Assigns a persona or expertise | Tone, depth, or perspective matters |
| System | Standing instructions for a whole session | Building an application or assistant |
| Instruction | A direct command | Almost always; it is the base of every prompt |
| Contextual | Supplies background or documents | The answer depends on facts the model lacks |
| Structured | Sections, delimiters, and a fixed layout | The prompt is long or machine-generated |
| Conversational | Multiple turns that build on each other | Exploring, refining, or clarifying |
| Multimodal | Images, audio, or files with text | The input is not only text |
| Retrieval-augmented | Context fetched at run time | Answers must come from your own data |
Zero-shot, one-shot, and few-shot prompting
These differ in the number of examples. Zero-shot relies on instructions alone. Few-shot shows the model input and output pairs before the real input.
Classify the review as positive or negative.
Review: "Arrived late and the box was damaged." -> negative
Review: "Does exactly what it says." -> positive
Review: "{{review}}" ->
Full guide: zero-shot, one-shot, and few-shot prompting.
Role prompting
A role prompt tells the model who to be: “You are a senior security reviewer.” It sets vocabulary, depth, and priorities in a few words. A role works best when it is specific and paired with a task; a role alone does nothing.
System prompts
A system prompt is the standing instruction an application sends before any user message. It defines the assistant’s job, rules, and output conventions for the entire conversation. Full guide: system prompts and role prompting.
Instruction prompting
The plainest type: a direct command, ideally one imperative sentence followed by requirements.
Rewrite the paragraph below at a grade 8 reading level. Keep every fact.
Every other type builds on a clear instruction. If the instruction is unclear, examples and roles will not rescue it.
Contextual prompting
Contextual prompts add the information the model needs to answer: a document, a product description, the audience, previous decisions. The skill is choosing what to include. Relevant context improves the answer; unrelated context competes with the instructions for attention.
Context: our refund window is 30 days and excludes custom orders.
Question: can a customer return a custom engraved item after 10 days?
Structured prompting
A structured prompt organizes its content into labeled sections, usually Task, Context, Requirements, Output format, and Input, with delimiters around variable data. It is the layout that scales: long prompts stay readable, and each part can be edited without disturbing the others. The PromptFlowEngine prompt optimizer can convert a paragraph into this layout and verify nothing was lost.
Do not confuse it with structured output, which is about the shape of the answer. See structured output prompting.
Conversational prompting
In a chat, each message is a prompt that includes the conversation so far. That lets you refine step by step: ask, read, correct, narrow. It is the fastest way to explore a problem.
Two cautions. Long conversations accumulate stale instructions that the model still tries to honor, so start a new conversation when the topic changes. And a result you reached through ten turns of correction is not reproducible until you write the final requirements down as a single prompt.
Multimodal prompting
Multimodal prompts combine text with images, audio, video, or documents. The same rules apply, with two additions: say what the model should look at (“in the chart on page 2”), and say what to do when something is unreadable, so it reports the problem instead of guessing.
The image is a photo of a paper receipt. Extract the merchant, the date,
and the total as JSON. If a field is not legible, use null.
Retrieval-augmented prompting
Here the application searches a knowledge base and inserts the results into the prompt before the question. The model answers from the supplied passages instead of from memory, which keeps answers current and lets it cite sources. Full guide: RAG prompting.
Choosing a type
Start with an instruction prompt. Add context if the model lacks facts. Add examples if the output is not what you want after the instructions are clear. Move it into a system prompt when it becomes part of an application. Add structure when it grows past a paragraph.
When simple shapes are not enough, the next step is technique: see prompt engineering techniques.