Prompt Engineering for Designers: Writing Instructions a Model Will Follow
Prompt engineering is designing the instructions, context and examples a model receives. What Anthropic, OpenAI and Google recommend, and a method to iterate.
What is a prompt?
A prompt is everything the model receives before it answers: role instructions, context, examples, history and the question. In many applications, much of that text is written by the team, not the user, and the user only contributes their part.
The field is young and its vocabulary is not standardized. The Prompt Report, a 2024 study, proposed a taxonomy of 58 text-based prompting techniques for language models and 33 vocabulary terms. Its authors point out that the area suffers from conflicting terminology. For a designer, that means focusing on a few well-understood techniques rather than an endless list.
What basic techniques do the providers recommend?
The official guides from Anthropic, OpenAI and Google agree on a common base:
- Clear, specific instruction. State the task, the audience and the goal. Anthropic suggests a test: show the prompt to a colleague with no context and ask them to follow it; if they get confused, the model will too. “Summarize this meeting for a manager who missed it” works better than “summarize this”.
- Context and motivation. Anthropic recommends explaining why a rule matters. Instead of “never use ellipses”, write that a text-to-speech engine will read the answer aloud and cannot pronounce them. Google and OpenAI also call for supplying supporting material.
- Examples. Two or three solved cases show the pattern (few-shot or multi-shot prompting). Anthropic suggests 3 to 5, relevant and varied. Google asks for specific examples with consistent formatting and warns that too many can cause overfitting. OpenAI asks you to cover a diverse range of inputs.
- Structure with delimiters. Anthropic uses XML tags to separate instructions, context and examples. OpenAI suggests Markdown for sections and XML to mark where a document starts and ends. Google mentions consistent delimiters.
- Output format. Say whether you want a table, a list or a paragraph. Google even proposes starting the answer yourself, for example with the first line of an outline.
- Role. Anthropic says that setting a role in the system prompt focuses behavior and tone, even with a single sentence.
Should you ask for step-by-step reasoning?
It depends on the model. The chain of thought technique, which makes the model write intermediate steps before answering, was described in 2022 (Jason Wei and others, at Google Brain). Today many models reason internally, and all three guides reflect that:
- OpenAI recommends giving high-level goals to reasoning models, not step-by-step instructions.
- Anthropic says an instruction such as “think thoroughly” often works better than a hand-written plan. If internal reasoning is switched off, asking for steps is still an option.
- Google states that its Gemini 2.5 and 3 models think internally, so you usually do not need to ask for the steps in the output.
With a model that has no built-in reasoning, asking for steps can still help. With one that reasons, it probably just spends tokens. Your evaluation decides, not habit. More in reasoning models.
Another technique from the field is self-consistency: generating several answers and keeping the conclusion that recurs most. It works when there is a verifiable answer, but it multiplies the cost.
What technique fits each case?
| Problem | Likely technique | Token cost |
|---|---|---|
| Format changes in every answer | Explicit output format and one example | Low |
| Tone is not the brand’s | Voice instruction with two approved examples | Low to medium |
| Mistakes in multi-step tasks | Clear goal and, depending on the model, reasoning | Medium, more output |
| Inconsistent answers on verifiable tasks | Self-consistency | High, several runs |
| The model lacks a fact | Not a prompt problem: use RAG | Depends on context |
How do you structure a product prompt?
A product prompt usually has five blocks: role, task, context, constraints and format. It is a variant of what the guides propose; OpenAI, for example, splits the message into identity, instructions, examples and context. Keeping the blocks separate makes it easier to see what changed:
- Role: who answers and for whom.
- Task: what it must produce, in one sentence.
- Context: the information the model needs and does not have. With long texts, Google suggests putting the material first and the instructions at the end.
- Constraints: what it must not do, with measurable limits.
- Format: the exact structure of the output.
How do you iterate a prompt?
Before iterating, Anthropic asks you to have your success criteria defined and a way to test them. Without that you cannot tell whether you improved anything. A method that works for design teams:
- Write the simplest prompt that describes the task.
- Gather about 20 real cases, including the difficult ones.
- Note what fails in each answer, with a category: format, tone, invented fact, omission.
- Change one thing at a time. If you change three instructions, you will not know which one improved the result.
- Run the cases again and compare with the previous version.
- Save the version with its date and the model used.
Prompts are sensitive to wording and to the order of content: Google lists rephrasing the prompt and changing the order of examples, context and input among its iteration strategies. OpenAI recommends pinning a specific model version in production, keeping prompts in code and reviewing them like any other change, and rerunning evaluations when the model is updated. For everyday conversational work, see iterating with prompts.
When NOT to write a long prompt?
When the problem is not about instructions. Anthropic notes that not every failing criterion is best solved with prompting: sometimes switching models improves latency and cost with less effort. If the model lacks the necessary information, a longer prompt only adds noise. If the task has exact rules, such as calculating a tax, use code. If you need a deep, repeated change of style, compare with fine-tuning vs. prompting. And if the prompt grows without control, it becomes hard to maintain and makes every request more expensive. Also check the context window so you do not fill it for no reason.
Example
Illustrative example: a team wants an assistant to write short descriptions for product cards. The first prompt says “write a description”. The results vary in length, some include prices and others use technical language. The second prompt defines the audience, a maximum of 140 characters, forbids prices and gives two approved examples. After 20 real cases, the team sees that tone still fails on premium products and adds a specific line for that case. Each change is tested separately and the final prompt is versioned.
Common mistakes
- Writing vague instructions and blaming the model.
- Changing several things at once and not knowing what worked.
- Testing only with the example the team had in mind.
- Giving nearly identical examples: Anthropic warns that the model may pick up patterns you did not intend.
- Dictating reasoning steps to a model that already reasons on its own.
- Keeping an unversioned prompt that nobody can reproduce later.
- Believing one perfect prompt works for any model. Each model change requires a review.