What are the main types of prompting?
Types of prompting are practical patterns for giving a model instructions and context. There is no single universal taxonomy. Common groups include zero-shot, one-shot, and few-shot prompting; role or persona instructions; retrieval-grounded prompts; structured-output prompts; multi-step workflows; and prompts that let a model call approved tools.
The labels describe how the task is framed, not a quality ladder. A zero-shot classification prompt can outperform a long example set when the categories are clear. Few-shot examples can help with edge cases, but they can also teach accidental patterns. Tool use matters when the answer requires current data or an external action rather than fluent prose.
Why types of prompting matter
Knowing the pattern helps a team choose the smallest amount of scaffolding needed for a reliable result. It also improves debugging. If the model misunderstands the output structure, add a schema or example. If it lacks facts, improve retrieval. If one request hides several decisions, split the workflow instead of making the instruction paragraph longer.
Classify prompts by the problem they solve: task definition, context supply, demonstration, decomposition, retrieval, tool use, or output control. Teams can combine patterns. A research workflow might retrieve approved sources, use a few examples to classify evidence, and return a fixed JSON object for a downstream system.
How to use types of prompting in practice
Use types of prompting inside a defined task rather than as an isolated piece of prose. Save the input, approved context, desired output, model, and test cases beside the prompt so later edits can be compared. Review a sample of real cases before relying on an aggregate score. The sample should include incomplete, conflicting, and high-value records because averages often hide the failures with the largest business cost. Place the documentation for types of prompting beside the workflow or decision it affects. That keeps definitions, evidence, ownership, and later corrections available to the people who need them.
Example
A content team asks a model to classify search queries by intent. The zero-shot version defines four labels and returns one label per query. After reviewers find confusion between comparison and purchase intent, the team adds five boundary examples. The prompt is now few-shot, while the task and output schema remain the same.
Choose a prompting pattern because it addresses a known failure. Naming a sophisticated technique does little if the workflow still lacks reliable evidence or a measurable definition of success.