Structure
Role
Role
Meaning
A description of the perspective or job the model should adopt.
When to use it
Use it to establish relevant expertise, audience, and tone.
Practical example
Act as a Korean technical editor for beginner developers.Generative AI Engineering
AI Prompt Design Terms
Learn how roles, instructions, context, examples, output schemas, multi-step workflows, and advanced reasoning patterns shape reliable AI responses.
18 matching terms
Structure
Instruction
Instruction
Meaning
A direct statement of the task the model should perform.
When to use it
Use an explicit verb and a clear deliverable to reduce ambiguity.
Practical example
Summarize the report into five decision-ready bullet points.Structure
Context
Context
Meaning
Background information the model needs to interpret the task correctly.
When to use it
Provide audience, goals, constraints, source material, and definitions that affect the answer.
Practical example
Context: the readers are first-time users and the feature launches next week.Structure
Constraint
Constraint
Meaning
A boundary on content, length, style, tools, or allowed actions.
When to use it
Use it to make acceptance criteria visible and testable.
Practical example
Use only the supplied policy, stay under 150 words, and do not infer missing dates.Structure
Delimiter
Delimiter
Meaning
A marker that separates instructions, examples, and source data.
When to use it
Use XML-like tags, headings, or fenced blocks when the prompt contains multiple sections.
Practical example
Summarize only the text inside <source>...</source>.Examples
Zero-shot prompting
Zero-shot prompting
Meaning
Requesting a task without providing a worked example.
When to use it
Start here for common tasks that the model already understands well.
Practical example
Classify each review as positive, neutral, or negative.Examples
Few-shot prompting
Few-shot prompting
Meaning
Providing a small set of input-output examples before the actual task.
When to use it
Use it to teach a specific label policy, tone, edge case, or output pattern.
Practical example
Input: delivery was fast. Output: logistics-positive. Now classify: the box arrived damaged.Reuse
Prompt template
Prompt template
Meaning
A reusable prompt with placeholders filled by variables at runtime.
When to use it
Use it for repeated workflows that require consistent instructions.
Practical example
Summarize {{document}} for {{audience}} using {{format}}.Output
Structured output
Structured output
Meaning
A response constrained to a predictable machine-readable shape.
When to use it
Use it when another program will parse or validate the result.
Practical example
Return an object with title, summary, risk_level, and source_ids.Output
JSON Schema
JSON Schema
Meaning
A formal description of allowed JSON fields, types, and required values.
When to use it
Use supported schema constraints to reduce parsing failures in automated workflows.
Practical example
Require status as an enum and due_date as a nullable ISO date.Workflow
Task decomposition
Task decomposition
Meaning
Breaking a complex request into smaller, verifiable subtasks.
When to use it
Use it when research, transformation, checking, and final synthesis require different steps.
Practical example
First extract requirements, then identify conflicts, then draft and verify the answer.Workflow
Prompt chaining
Prompt chaining
Meaning
Passing the output of one model step into a later model step.
When to use it
Use it when intermediate results should be inspected, transformed, or approved.
Caution
Errors can propagate across steps, so validate important intermediate outputs.
Practical example
Extract claims -> retrieve evidence -> verify each claim -> write the final brief.Advanced reasoning patterns
Step-back prompting
Step-back prompting
Meaning
Starting with a broader principle, definition, or decision framework before answering a specific question.
When to use it
Use it when the specific case is easy to misread without first identifying the governing concepts or constraints.
Caution
The broader step can still be wrong or irrelevant, so verify it against authoritative sources.
Practical example
First identify the policy principles that govern refund eligibility. Then apply them to this case and cite the supporting clauses.Advanced reasoning patterns
Least-to-most prompting
Least-to-most prompting
Meaning
Solving simpler dependent subproblems first and using their verified results to address the harder overall task.
When to use it
Use it when later conclusions depend on several prerequisite calculations, facts, or decisions.
Caution
An early error can propagate through later steps; validate prerequisite outputs before reuse.
Practical example
1. Extract the contract dates. 2. Calculate each notice period. 3. Compare the periods with the termination clause. 4. Report the conclusion with evidence.Advanced reasoning patterns
Self-consistency
Self-consistency
Meaning
Generating multiple independently sampled candidate solutions and selecting an answer supported by agreement or a separate verification rule.
When to use it
Use it selectively for difficult tasks where multiple valid solution paths can be compared and the additional cost is justified.
Caution
Agreement does not prove correctness, correlated errors can agree, and multiple samples increase latency and cost.
Practical example
Generate five independent candidate totals, compare the final values, and accept one only if the source rows and deterministic calculation verify it.Advanced reasoning patterns
Critique-and-revise
Critique-and-revise
Meaning
Producing a draft, evaluating it against explicit criteria and evidence, and revising the identified weaknesses.
When to use it
Use it when the output has a reviewable rubric such as completeness, source support, format, tone, or policy compliance.
Caution
A model can overlook its own mistakes; high-impact results still need independent checks or human review.
Practical example
Draft the answer. Critique it against required facts, citations, prohibited claims, and length. Revise only the failed items, then return the final answer.Advanced reasoning patterns
Plan-and-execute
Plan-and-execute
Meaning
Separating task planning from execution so dependencies, tools, checkpoints, and completion criteria are defined before work proceeds.
When to use it
Use it for multi-step work where order, permissions, resource limits, or intermediate validation matter.
Caution
Treat the plan as revisable; do not continue blindly when new evidence invalidates it.
Practical example
Return a concise action plan with dependencies and approval points. Execute each approved step, record evidence, and replan when an assumption fails.Advanced reasoning patterns
Verifier pattern
Verifier pattern
Meaning
Using a distinct verification step to check a candidate answer against rules, evidence, schemas, or deterministic tools before acceptance.
When to use it
Use it when important claims, calculations, citations, structured outputs, or tool arguments can be checked independently.
Caution
A verifier that shares the same missing evidence or flawed assumptions may repeat the generator's error.
Practical example
Generator: produce the invoice summary with source row IDs. Verifier: recompute totals, validate the schema, and reject unsupported claims.Choosing an advanced reasoning pattern
Select a pattern only when the task structure and evaluation evidence justify the additional steps, tokens, latency, and complexity.
| Need | Pattern to consider |
|---|---|
| Identify governing principles before applying them to a case | Step-back prompting |
| Solve prerequisite subproblems in dependency order | Least-to-most prompting |
| Compare several candidate solutions when additional sampling is justified | Self-consistency plus independent verification |
| Improve a draft against explicit review criteria | Critique-and-revise |
| Coordinate dependencies, tools, and approval checkpoints | Plan-and-execute |
| Reject unsupported claims, invalid schemas, or wrong calculations | Verifier pattern |
Frequently asked questions
Should a prompt ask the model to reveal its complete internal chain of thought?
No. Prefer a concise answer, checkable intermediate results, cited evidence, assumptions, calculations, and validation outcomes. Some reasoning models perform internal thinking automatically, and a long visible rationale is not proof that the answer is correct.
Do advanced reasoning patterns always improve accuracy?
No. Their effect varies by model, task, prompt, sampling, and evaluation method. They can also add latency, token cost, correlated errors, and failure propagation. Compare them with a simple baseline on representative cases.
How is a verifier different from critique-and-revise?
Critique-and-revise improves a draft through review and editing. A verifier acts as an acceptance gate that checks a candidate against evidence, rules, schemas, or deterministic tools and may reject it without rewriting it.
Official references
- Google AI for Developers — Prompt design strategies
Official guidance on structured prompting, reasoning, planning, validation, agentic workflows, and model-specific thinking behavior.
- Wang et al. — Self-Consistency Improves Chain of Thought Reasoning in Language Models
Primary research introducing self-consistency as sampling diverse reasoning paths and selecting the most consistent answer.
Related terminology guides
- AI Context Engineering Terms
Extend a prompt into a controlled context containing evidence, history, state, and reusable material.
- AI Multimodal Prompt Engineering Terms
Design requests that combine text with images, documents, audio, or video.
- AI Prompt Testing and Optimization Terms
Compare prompt variants with repeatable cases, metrics, and regression checks.
- AI Prompt Management and Operations Terms
Version, approve, deploy, observe, roll back, and retire production prompts.
- RAG and Document Search Terms
Apply query rewriting, evidence-only answers, abstention, citations, synthesis, and source-conflict handling.