More products now ask people to tell an AI what they want. That sounds simple until you watch someone face an empty chat box. The user may know the outcome they need, but not the language, constraints or examples that will get them there. That gap is a UX problem.

I see prompt design as a first-class UX skill alongside wireframing, interaction design and UX writing. It is not about discovering a magic phrase. It is about designing an interaction that helps a person express intent, understand what the system can do and recover gracefully when it cannot.

Why the wireframe analogy holds up

A prompt is an interface specification in miniature. Like a wireframe, it makes assumptions visible, needs edge cases, and improves through quick rounds of testing. The output may be language rather than pixels, but the design work is familiar:

• Clarify the user’s goal before asking for detail.

• Make constraints visible rather than forcing people to remember them.

• Test realistic tasks with real people, not only the team that wrote the prompt.

• Design the failure state: uncertainty, unsafe requests and missing information are normal, not embarrassing exceptions.

The prompt pattern I reach for

For internal tools and prototype work, I use a simple structure: Context → Task → Constraints → Output format → Example. It stops the model from filling in the blanks with misplaced confidence.

Context: You are a professional UX writer with 8+ years of experience.
Task: Turn the following feature description into a 3-bullet microcopy for onboarding.
Constraints: 2–3 sentences per bullet, friendly tone, no technical jargon.
Output format: JSON array of strings.
Input: {FEATURE_DESCRIPTION}
Example:
["Welcome to X — ...", "Tip two — ..."]

Context: You are a professional UX writer with 8+ years of experience.
Task: Turn the following feature description into a 3-bullet microcopy for onboarding.
Constraints: 2–3 sentences per bullet, friendly tone, no technical jargon.
Output format: JSON array of strings.
Input: {FEATURE_DESCRIPTION}
Example:
["Welcome to X — ...", "Tip two — ..."]

Context: You are a professional UX writer with 8+ years of experience.
Task: Turn the following feature description into a 3-bullet microcopy for onboarding.
Constraints: 2–3 sentences per bullet, friendly tone, no technical jargon.
Output format: JSON array of strings.
Input: {FEATURE_DESCRIPTION}
Example:
["Welcome to X — ...", "Tip two — ..."]

Context: You are a professional UX writer with 8+ years of experience.
Task: Turn the following feature description into a 3-bullet microcopy for onboarding.
Constraints: 2–3 sentences per bullet, friendly tone, no technical jargon.
Output format: JSON array of strings.
Input: {FEATURE_DESCRIPTION}
Example:
["Welcome to X — ...", "Tip two — ..."]

It is deliberately unglamorous. Good UX often is. The point is to produce something a team can review, version and improve—not a one-off incantation that only works while its author is in the room.

Design the interface around the prompt

The best AI experiences do not make every customer become a prompt engineer. They provide helpful scaffolding:

• Prompt builders for audience, tone, length and source material.

• Examples and style choices that show what a good request looks like.

• Editable assumptions so the user can correct the system before it runs away with them.

• Preview and revision controls that make iteration feel safe and legible.

These patterns reduce what researchers call the articulation barrier: the difficulty people have turning a fuzzy need into a precise request. The interface should carry some of that cognitive load. A blank box is not a strategy.

How I have used this in discovery

At ShipServ, I used AI-assisted prompting as part of early exploration for an e-invoicing concept. We created low-fidelity interactive concepts quickly, then put them in front of users with Maze before committing to a build. The useful unit of work was not “the prompt”; it was the whole loop: prompt, prototype, task, observation, revision. I cover that process in more detail in my AI-assisted prototyping post and the ShipServ case studies.

A short checklist before shipping

1. Can a first-time user tell what the assistant is for?

2. Does the product collect the minimum context needed for a useful answer?

3. Can the user see, edit or reject important assumptions?

4. Have you tested messy, ambiguous and incomplete inputs?

5. Is there a clear human next step when the model is uncertain?

Prompt design will evolve as models and interfaces do. The durable UX skill is not memorising syntax; it is helping people get a trustworthy result with the least unnecessary effort. That is a very old design problem wearing a new, slightly overconfident, hat.