The 5 Building Blocks Every Prompt Needs

The 5 building blocks every prompt needs are: Role & Context, Task / Instruction, Input & Examples, Constraints, and Output Format. These five components are the minimum structure that separates a reliable, repeatable prompt from a vague question that produces inconsistent results.

Each block solves a different failure mode. Role & Context tells the model who it is and what situation it is in. Task / Instruction tells it exactly what to do. Input & Examples give it the raw material and teaching signal. Constraints set the rules. Output Format specifies the shape of the answer. Together, they leave the model with nothing left to guess.

The five-block model reflects the converged consensus across prompt engineering guidance from OpenAI, Google, Anthropic, and independent practitioners. Role, instructions, examples, constraints, and output format appear — under different names — in every major framework published since 2023. This is not coincidence: it is the minimum information a probabilistic model needs to produce a useful, consistent result.

The business case is straightforward. Missing role and context produces generic answers that need rewriting. Missing constraints increases hallucination risk and off-brand output. Missing output format means results that cannot be parsed or copy-pasted directly. The 5-block model addresses all three failure modes at once, and applies equally to GPT-4-class models, Claude, Gemini, and locally-run LLMs.

Role tells the model what persona or expertise to adopt. Context tells it the situation, domain, and audience it is operating in. They are grouped together because they work as a pair — role is who the model is, and context is the environment that shapes what "good" means for that task.

When you omit role and context, the model answers from a generic perspective — useful to no one in particular. With them, the same model becomes a senior tax advisor answering a question about VAT returns, a junior copywriter writing for a 19-year-old audience, or a data analyst summarising a quarterly report. The output calibrates to your actual situation.

To understand how role prompting evolved as models became more capable of following instructions, see From GPT-2 to Today: How Prompt Engineering Evolved.

The Task / Instruction block is the explicit statement of what you want the model to do. It is the most important block — every other block supports this one. A clear, specific, testable instruction reduces ambiguity to near-zero. A vague instruction is the single biggest cause of poor AI output across all models and use cases.

Current best-practice guidance emphasises making the task actionable and observable: use a verb, state the deliverable, and where possible describe a success criterion. "Write a summary" is a task. "Summarise the following article in 3 bullet points, each under 20 words" is a task with a testable output. The difference in output quality is significant.

Instructions interact directly with whether you provide examples — see Zero-Shot vs. Few-Shot: Which Approach Gets Better Results? for when each approach works best.

Input is the actual data, text, or material the model needs to work on. Examples are sample input/output pairs that demonstrate what a correct response looks like. These are separate concerns: input is the raw material for the current task, examples are the teaching signal that shapes how the model performs it.

Including 1–3 examples (few-shot prompting) is the single most reliable technique for locking in output format and tone. When you show the model what a good answer looks like, it matches the pattern rather than inferring it from the task description alone. This matters most for specialised formats, consistent tone, and structured outputs where precision is required.

For a full breakdown of when to use examples and when to omit them, see Zero-Shot vs. Few-Shot: Which Approach Gets Better Results?.

Constraints are the rules the model must follow: what it must do and what it must not do. They include length limits, forbidden topics or phrases, required sources, brand voice rules, safety boundaries, and format restrictions. Constraints are the most commonly omitted block — and their absence is the primary cause of hallucinated facts, off-brand language, and outputs that arrive in the wrong format.

Adding one well-scoped constraint is often the highest-leverage change you can make to an existing prompt. "Do not make up statistics" cuts hallucination risk sharply. "Never exceed 100 words" forces concision. "Only use information from the text provided" grounds the output in the source material and eliminates fabrication entirely for that task.

For the technique of using exclusions to shape output, see Negative Prompting: Tell the AI What NOT to Do. For why unconstrained prompts hallucinate more, see AI Hallucinations: Why AI Makes Things Up — and How to Stop Them.

Output Format specifies the exact shape of the answer the model should produce. This is the block that determines whether the output is directly usable or requires reformatting before it is useful. For automated pipelines, an unspecified output format means brittle, inconsistent parsing. For GEO, a structured output is more likely to be cited verbatim by AI search engines, because structured answers are easier to extract programmatically.

The output format block can specify the file format (JSON, Markdown, CSV), the structure (table, bullet list, numbered steps), the length, and the labelling of sections. The more precisely you specify it, the less editing the output requires.

For a full guide to JSON mode and structured output across different models, see Structured Output & JSON Mode: Get AI to Return Usable Data.

The template below shows all 5 blocks assembled in order for a single domain-neutral task. Each part is labelled so you can see exactly where each block begins and ends. Replace the content in each section to adapt it to any domain.

This template works on GPT-4o, Claude, Gemini, and local LLMs via Ollama or LM Studio. The block order is a recommendation, not a rigid rule — but placing Role & Context first and Output Format last is the most common and reliable arrangement across all major models.

Popular prompt engineering frameworks are opinionated ways to arrange the same five building blocks under different names and in different orders. CRAFT, CO-STAR, and SPECS all map directly to this five-block model. Understanding the blocks first means you can apply any framework without memorising its specific terminology from scratch.

The table below shows how each building block maps to the corresponding field in three widely used frameworks:

PromptQuorum includes 9 built-in frameworks that pre-fill these blocks in different configurations depending on the task type. For framework-specific guides, see Which Prompt Framework Should You Use?, CRAFT Framework, and CO-STAR Framework.