Prompting is product design
By Imran Sheikh · 2026-11-12 · AI in practice · 4 min read
Key points
- A system prompt sets tone, limits and behavior for every user.
- Write prompts like specs, then test them against real cases.
- Designers belong in the prompt, because words are the interface.
In an AI product, the system prompt is part of the interface. It decides how the product talks, what it will do and where it stops. Every person who uses the product meets it, even though no one sees it.
That makes prompting a product design task. Designers already know how to decide on voice, behavior and boundaries. Those skills apply directly. The medium is text, and text is something many designers are good at.
The prompt is a spec that runs
A product spec describes what a feature should do. A system prompt does the same, and then the model acts on it. A vague spec gives vague behavior. A contradictory spec gives erratic behavior. You get what you wrote, not what you meant.
This is a gift to designers. You can change the product by changing words, and you can see the effect in minutes. You can compare two versions side by side. The feedback loop is far shorter than in most design work.
What a good prompt contains
Strong prompts tend to share a structure. They begin with a role and a purpose. They describe the person being served. They set the tone with examples. They state what to do when the request is unclear or out of scope.
- Purpose: what the product is for and who it serves.
- Voice: how it sounds, with a few sample lines.
- Behavior: how long to answer, when to ask a question, when to stop.
- Boundaries: what it will not do, and how to decline kindly.
- Format: how to present lists, steps and warnings.
- Fallbacks: what to do when it does not know.
Notice how much of this is the content of a design brief. Voice, tone, edge cases and error behavior are familiar ground. A designer who has written a style guide already knows how to write half of a system prompt.
Show, do not only tell
Examples beat adjectives. Telling a model to be warm and concise gets a generic result. Showing three sample replies in the right voice gets something much closer. Pick examples that cover the hard cases, not only the easy ones.
Include an example of a good refusal. Include an example of an honest I do not know. Include a short answer to a long question. The model learns the shape of good behavior from these, and so does the next person who reads the prompt.
Test like a designer
A prompt is not done when it reads well. It is done when it survives real use. Build a set of test inputs and run every revision against them. Keep the set growing as you meet new failures.
- Typical requests from your actual users.
- Vague requests that need a clarifying question.
- Requests outside the product’s scope.
- Emotional or sensitive messages.
- Attempts to pull the product off its rules.
- Very short and very long inputs.
Read the outputs yourself. Give each output a full read. Look for tone drift, needless caveats and answers that are technically right and humanly cold. Mark what you would change and trace it back to a line in the prompt.
Fewer rules, clearer intent
A common mistake is the endless list of rules. Every failure adds another line, until the prompt is a thicket that no one can reason about. Rules collide. The model follows some and drops others.
Prefer clear intent to long lists. Explain why a behavior matters, and the model can handle cases you did not foresee. Keep a short set of firm rules for the things that must never happen. Review the prompt regularly and delete lines that no longer earn their place.
Version it like code
Keep prompts in the repo, next to the code they affect. Review changes in pull requests. Write a short note on why each change was made. This gives you history, and it lets designers and engineers work on the same file.
A prompt also needs an owner. Without one, every teammate adds a line to fix their own complaint, and the voice slowly turns to mush. Name a person who guards the tone and the structure. Others can propose edits, and the owner decides which ones stay.
It also fits how we work at Emerge Labs. We design in the client’s codebase, and for AI features the prompt is part of that code. A designer who can edit a prompt, run the tests and open a pull request can shape the product directly.
In a chat product, nearly everything the person experiences is language. The greeting, the questions, the confirmations, the errors, the goodbye. All of it comes from the prompt and the model working together. If you care about the experience, you have to care about those words.
We build our own AI products at the studio and the lesson repeats. The best improvements often come from a single sentence in the prompt. A better instruction about when to ask a question can change the whole feel of a conversation.
If words are the interface, the prompt is the design file.
One more habit helps. Read the prompt aloud as if you were the person using the product. If a line sounds odd or cold when spoken, it will sound odd or cold in the product too. Fix the sentence and run the tests again.
So put designers in the prompt. Give them access, give them a test set and give them time. The result is a product that sounds like it was made by someone who cared, because it was.