Prompting academy
  • Level: foundations
  • 9 min read

Prompt formats: how to write an instruction that gives repeatable results

The length of a prompt does not decide the quality of the answer. The structure does. Below are five proven formats that organise an instruction so an agent answers just as well on the first attempt as on the hundredth.

A prompt is not a question but a specification of a task. The more clearly you describe the role, the context and the expected result, the less room you leave for chance. The five formats below work as scaffolding: pick one, fill it with your own content and get a repeatable result.

Why format matters more than length

Language models do not read minds. When an instruction is one long run of sentences, the model decides for itself what is the goal and what is only an example. When you separate those elements, you stop guessing why it works one day and not the next.

  • Role and goal — who the agent is and what it must achieve.
  • Context — facts, data and limits it must respect.
  • Output format — what the answer should look like.
  • Examples — a model of a good and a bad result.
  • Prohibitions — what the agent must never do.

SPICE — structure for longer tasks

SPICE organises long instructions into five sections. It works wherever a prompt runs to several pages and describes a whole process rather than a single question.

  • Sections — divide the instruction into clear blocks.
  • Prompt variables — places you swap with each use.
  • Instructions — the steps the agent must perform.
  • Context — the knowledge and data it works from.
  • Examples — models of a correct result.

When to use SPICE

When you build a fixed conversation scenario, a product description or a support procedure the agent must always follow the same way. Sections let you return to the prompt weeks later and immediately see what is responsible for what.

DELTA — designing a role and its behaviour

DELTA answers how the agent should behave, not just what it should do. It is the most effective format for designing roles, personas and tone of voice.

  • Describe — describe the agent's goal and task.
  • Example — show examples of the style you expect.
  • Limitations — list prohibitions and boundaries.
  • Tone — define the style of communication.
  • Audience — describe who the agent speaks to.

When to use DELTA

Whenever the agent should sound like a specific person: an adviser, a consultant, a customer care specialist. DELTA is also the basis of meta-prompting, that is writing a prompt that creates another prompt.

Tabula Rasa — theory first, task second

Tabula Rasa reverses the order. Instead of assigning the task straight away, you first ask the agent to organise knowledge: theory, good practice, examples. Only then do you assign the actual work.

The effect: the agent loads the right context before answering, and the result is far more precise. This technique is especially useful in conversations where the customer asks questions from many fields.

Prompt Chaining — splitting a complex task

Prompt Chaining breaks a large task into smaller steps. The output of one step becomes the input of the next, so nothing gets lost along the way.

  • Step 1: collect information about the customer's need.
  • Step 2: organise it into a set of parameters.
  • Step 3: assess completeness and flag the gaps.
  • Step 4: prepare a summary and the next step.

In practice the chain turns a chaotic conversation into a predictable process. Each stage can be checked and corrected on its own, without rewriting the whole thing.

Chain of Thought — reasoning step by step

Chain of Thought means the agent first lays out its reasoning and only then gives the answer. That makes it easier to spot where a mistake was made.

This approach increases precision in tasks with many conditions: pricing, enquiry qualification, choosing a variant. You see not only the answer but also the path to it.

How to combine the formats into one scenario

The formats do not exclude each other. Most often they are combined into a simple sequence: SPICE organises the whole, DELTA describes the role, Tabula Rasa loads the knowledge, Prompt Chaining divides the work and Chain of Thought guards the logic.

  • Start with SPICE — give the prompt structure.
  • Add DELTA — define role, tone and audience.
  • Turn on Tabula Rasa — let the agent organise knowledge.
  • Divide the task with a chain — step by step.
  • Close the logic with Chain of Thought — check the reasoning.

You paste the finished prompt into ePersona, where it becomes a conversation scenario with your customer. The Academy teaches how to write it — the tool tests it in a real conversation.

Frequently asked questions

Where should I start learning prompting?

With one format. SPICE is best to begin with, because it organises the whole instruction; role, examples and prohibitions come later.

Is a longer prompt better?

No. A well-organised prompt is better. Sections, examples and clear prohibitions give a repeatable result more reliably than length alone.

What is the difference between a system prompt and a normal question?

A system prompt describes the agent's role, rules and way of working permanently. A normal question concerns one answer.