Most designers use AI reactively.

They open ChatGPT when:

• they do not know where to begin,
• they need to write documentation quickly,
• they are stuck between two versions of a solution,
• the Product Manager wants a summary “by the end of today”,
• the meeting notes resemble a transcript of five people who each attended a different meeting,
• their brain has already finished work, but the sprint apparently has not received the message.

Then comes the quick prompt:

Improve this flow.

Or:

Write a professional design rationale.

AI responds. The text looks correct. It is structured, confident, and contains several words such as seamless, user-centred, and enhance. In other words, it looks exactly like the response the same model could prepare for ten thousand other designers.

The problem is not only a weak prompt.

The problem is the lack of a process.

Working effectively with AI is not about collecting increasingly elaborate instructions. It is about consciously managing the entire collaboration:

• what you do before asking a question,
• what context you provide,
• how you use the response,
• how you evaluate it,
• what you remove,
• what you add,
• which parts remain your responsibility.

In 15 UX Prompts for ChatGPT That Support Real Design Work, I shared specific prompts for research, analysis, documentation, edge cases, and portfolio work.

Here, we will focus on something more important:

How to build a workflow in which AI supports your thinking without taking control of its direction.

An AI Workflow Is Not a List of Prompts

A prompt is a single instruction.

A workflow is a system that determines:

• how you define the goal,
• how you provide context,
• how you expand the range of possibilities,
• how you select the useful material,
• how you verify its accuracy,
• how you add your own judgement,
• how you turn the result into a decision or usable deliverable.

You can have an excellent prompt and still receive a weak result if:

• you do not know what you actually need,
• you have not provided constraints,
• you treat generated possibilities as validated solutions,
• you accept the first answer,
• you do not verify the facts,
• you do not understand the consequences of the recommendation,
• you allow AI to add experience that never existed.

A simple prompt, on the other hand, can be extremely useful when it forms part of a well-managed process.

In The Designer’s AI Playbook, I propose a four-stage model:

Lead → Expand → Refine → Elevate

Each stage has a different purpose.

Lead

You set the direction.

Expand

AI broadens the range of possibilities.

Refine

You select, verify, and organise the material.

Elevate

You add judgement, experience, context, and responsibility.

Most weak results appear because the user starts with the second stage.

They ask AI to generate before deciding:

• what the problem is,
• what they want to learn,
• which constraints exist,
• which information is reliable,
• which criteria they will use to evaluate the answer.

AI receives an empty stage, so it does what it does best: it produces possibilities.

Sometimes very convincing ones.

Not always useful ones.

1. Lead — Set the Direction

Your first task is not to write a long prompt.

Your first task is to establish where you are actually trying to go.

The Lead stage should include four elements:

1. context,
2. intent,
3. constraints,
4. perspective.

Context: What Situation Are You Working In?

AI does not know your project simply because you have spent the past three months working on it.

Explain:

• what kind of product it is,
• who the users are,
• which process you are analysing,
• what stage the project has reached,
• which roles or systems are involved,
• what information you already have.

Instead of:

Analyse this booking flow.

Write:

I am designing a university equipment-booking process. A student searches for a device, checks the requirements, selects a date, and submits a reservation. Some equipment can be booked automatically, while specialist devices require staff approval.

The second version is not more effective because it is longer.

It is more effective because it reduces the space available for guessing.

Intent: What Do You Need?

Do you want to:

• find missing scenarios,
• compare two solutions,
• prepare a flow description,
• organise notes,
• write a stakeholder summary,
• prepare questions for a developer,
• test the logic of a decision,
• create a case study structure?

“Help me with this flow” does not define an outcome.

This is much clearer:

I want to identify the situations in which a user cannot complete the reservation and verify whether the system explains the reason for the restriction clearly enough.

Now the purpose is clear.

Constraints: What Should AI Not Do?

This is one of the most important parts of a strong workflow.

You can specify:

• do not add new features,
• do not assume integrations that have not been described,
• do not invent data,
• do not choose the solution for me,
• do not present hypotheses as facts,
• preserve the product terminology,
• do not change the business rules,
• separate user behaviour from system behaviour,
• identify missing information instead of filling the gaps.

Example:

Do not assume that the system can confirm every reservation automatically. Do not add new roles or features. If information is missing, record it as a question that requires clarification.

Without that restriction, AI may enthusiastically suggest:

• an intelligent recommendation system,
• availability prediction,
• approval automation,
• a chatbot,
• dynamic notifications,
• and probably a dashboard, because there is always room for a dashboard somewhere.

Perspective: How Do You Currently Understand the Problem?

AI should not replace your point of view, but it can challenge it.

State your current assessment:

I believe the main problem is not the number of steps, but the fact that users learn about the required permissions too late.

Or:

My hypothesis is that manually approving every reservation creates unnecessary staff workload, but we still need control over exceptional cases.

The model no longer begins the conversation from nothing.

It can:

• expand your hypothesis,
• identify its weaknesses,
• offer another perspective,
• help clarify what you still do not know.

A Complete Lead Stage

I am designing a university equipment-booking process. A student searches for a device, checks the requirements, selects a date, and submits a reservation. Some devices require completed training, while unusual cases are reviewed by a staff member.

I want to verify whether the flow handles restrictions, exceptions, and interrupted processes appropriately.

Do not add new features or roles. Do not assume the system has access to information I have not provided. Separate confirmed rules from questions that require clarification.

My current hypothesis is that permission requirements appear too late, which means users may begin a reservation they cannot complete.

This is not a spell.

It is a properly prepared brief.

2. Expand — Broaden the Possibilities

Only now is it time for AI to do what it is genuinely good at:

• generating variations,
• identifying patterns,
• proposing alternatives,
• finding edge cases,
• simulating different perspectives,
• breaking a problem into smaller parts,
• comparing possible consequences.

The Expand stage is not used to make the decision.

It is used to increase the number of things you can notice before making it.

That distinction matters.

You are not collecting answers. You are collecting possibilities.

Useful Questions During the Expand Stage

• What alternative solutions could we consider?
• How might this process work for different roles?
• Which failure scenarios may be missing?
• What happens if the user interrupts the process?
• What are the possible consequences of removing this step?
• Which conflicting needs exist within this process?
• What questions might a developer, analyst, or compliance officer ask?
• Which user behaviours could challenge the happy path?
• Which criteria would be useful when comparing the solutions?

For our example:

Generate possible scenarios involving missing permissions, expired training, availability changing during the reservation, interrupted processes, date conflicts, delayed data, and manual review of exceptional cases.

For each scenario, identify:

• when it occurs,
• the possible consequence,
• the information the user needs,
• a question that requires confirmation from the team.

AI may generate twenty scenarios.

That does not mean all of them are probable, important, or technically possible.

You do not evaluate them yet.

First, you broaden the map.

Expand Does Not Mean “Generate the Entire Product for Me”

There is a difference between exploration and delegation.

Exploration:

Show me three ways of handling reservations that require approval and compare their consequences.

Delegation:

Design the best booking flow.

The first instruction helps you see the options.

The second pretends that AI can choose the solution without fully understanding the users, the system, the costs, the organisation’s policies, and the business responsibility.

It cannot.

It can generate the most common flow.

Common does not mean appropriate.

Use Different Perspectives, but Do Not Pretend They Are Research

During the Expand stage, you can ask AI to analyse the problem from the perspective of:

• a new user,
• an expert user,
• an administrator,
• a developer,
• a security specialist,
• a business stakeholder.

Example:

Analyse this decision from the perspectives of a student, the staff member issuing the equipment, and the developer maintaining the system. For each role, identify possible benefits, costs, and risks.

This may reveal questions you had not previously considered.

It does not mean that AI represents a real student, staff member, or developer.

It is a simulated perspective.

Not evidence.

If the model states:

Students prefer immediate confirmation.

you have not received a research finding.

You have received a statement that needs to be tested.

3. Refine — Narrow, Verify, and Organise

The Expand stage may give you:

• many variations,
• a dozen scenarios,
• a list of risks,
• alternative structures,
• several ways of framing the problem.

Now the most important work begins.

Refine means:

• rejecting,
• combining,
• simplifying,
• verifying,
• prioritising,
• correcting the logic,
• restoring alignment with the real product.

This is where the material begins to become yours.

AI can generate a draft.

It does not know:

• which assumption conflicts with a client conversation,
• which feature is outside the scope,
• which scenario is technically impossible,
• which term has a different meaning within your product,
• which risk is unacceptable from a business perspective,
• which solution conflicts with organisational policy.

Questions for the Refine Stage

Review the response and ask:

What Is Based on the Context I Provided?

Separate information derived from the material from generated assumptions.

What Requires Confirmation?

Do not allow the model to fill gaps with elegant prose.

What Is Outside the Scope?

Remove features and scenarios the project does not cover.

What Is Repetitive?

AI is very fond of saying the same thing four times, each time in slightly more “strategic” language.

What Actually Matters?

Not every edge case deserves the same priority.

What Changes the Decision?

Keep information that may affect the flow, architecture, message, or scope.

Can I Understand and Defend Every Recommendation?

If not, the answer is not ready to use.

Example Refine Prompt

Remove scenarios that require features not available in the described system. Combine similar cases.

Divide the remaining items into:

• critical to completing the reservation,
• important but suitable for later handling,
• requiring confirmation from the team.

Do not propose the interface yet. Preserve the terminology: “reservation”, “eligibility”, “equipment readiness”, and “manual review”.

Notice the final sentence.

Controlling terminology is part of product design.

If you use approval once, verification later, and then confirmation, the interface may end up describing three different processes with one chaotic vocabulary.

AI will not correct that unless you define what the product objects and statuses mean.

Refine Also Requires Verification Outside AI

Not everything should be resolved with another prompt.

Depending on the material, you may need:

• a conversation with a developer,
• verification of the business requirements,
• consultation with an analyst,
• a return to the research,
• input from a security specialist,
• comparison with the behaviour of the existing system,
• testing with users.

AI may identify the question:

What happens if the user’s training expires after the reservation is made but before the equipment is collected?

The answer should not be invented by the model.

It is either a product decision or an existing business rule.

4. Elevate — Add Human Judgement

After the Refine stage, you may have correct material:

• an organised analysis,
• a document structure,
• a comparison of options,
• a flow description,
• a list of decisions,
• an initial version of the message.

Correct does not necessarily mean valuable.

The Elevate stage adds what AI cannot provide independently:

• interpretation,
• experience,
• hierarchy of importance,
• knowledge of the organisation,
• awareness of risk,
• point of view,
• recommendation,
• responsibility.

This is where you move from:

Option C balances automation and control.

to:

I chose conditional approval because most standard reservations follow repeatable rules and do not require a staff decision. Manual review remains necessary for specialist equipment, expired permissions, and exceptional cases. The solution shortens the standard path but increases the importance of accurately defined eligibility logic.

The second version is not simply longer.

It shows:

• what you chose,
• why,
• what you gained,
• what cost the decision introduced,
• where the risk lies.

That is the design layer.

Elevate Does Not Mean “Make It Sound More Senior”

One of the most misleading prompts is:

Make it sound more senior.

AI will probably add:

• strategic alignment,
• scalable framework,
• holistic approach,
• seamless experience,
• business value.

The text will begin to sound more expensive.

Not necessarily more intelligent.

Seniority does not come from tone. It comes from the quality of the reasoning.

Instead of requesting “senior language”, add:

• the consequences of the decision,
• trade-offs,
• risk,
• dependencies,
• selection criteria,
• unresolved questions,
• impact on the user and the system.

Example:

Add the consequences of the decision for the user, staff workload, and system-logic complexity. Do not add business impact that has not been confirmed.

The text becomes more mature because it contains more meaningful information.

Not because it put on a LinkedIn blazer.

A Complete Example: Deciding How Reservations Should Be Approved

Let us examine the entire workflow in one process.

Lead

I am designing a university equipment-booking system. At present, every reservation requires manual staff approval. Most basic equipment follows the same rules, but specialist devices require training and additional verification.

I want to compare three possible approval models.

Do not assume changes to the university’s policies or new integrations. Do not choose the solution for me. Identify the trade-offs.

Expand

Compare:

1. manual approval of every reservation,
2. automatic approval of every reservation,
3. automatic approval of standard reservations with manual handling of exceptions.

Evaluate them in terms of user time, staff workload, risk of error, clarity of statuses, and system complexity.

AI prepares a broad comparison.

Refine

Remove arguments that assume features not described in the context.

Separate the risks into:

• user-related,
• operational,
• technical.

Identify which rules must be defined for conditional approval to be possible.

You now have material that is more closely aligned with the real problem.

Then you verify it with the team.

You may learn, for example, that the system contains training-completion data but updates it only once per day. This creates a new risk.

Elevate

Based on the verified information, you formulate the decision:

I chose conditional approval. Standard loans are confirmed automatically when the user meets the requirements and has no overdue returns. Specialist-equipment reservations, outdated training data, and exceptional cases are sent to a staff member.

The solution reduces the number of manually reviewed requests, but it depends on the accuracy of the permissions data. Eligibility status therefore needs to be visible before the user selects a date, rather than only after they submit the form.

AI helped you:

• generate options,
• break down the criteria,
• notice risks,
• organise the material.

It did not make the decision.

It did not know about the system’s update schedule.

It is not responsible for the consequences.

You are.

How to Use the Workflow Across Different Areas of Product Design

The Lead → Expand → Refine → Elevate model is not limited to flow analysis.

It can organise many forms of design work.

Research

Lead

Provide:

• the research objective,
• the available material,
• the questions,
• the sample limitations.

Expand

Ask for:

• possible themes,
• contradictions,
• missing questions,
• alternative interpretations.

Refine

Check:

• which conclusions are supported by evidence,
• which rely on a single statement,
• what cannot be concluded,
• what requires further research.

Elevate

Add:

• the context of the conversations,
• your knowledge of the users,
• relevance to product decisions,
• your interpretation and confidence level.

AI can organise data.

It should not create insights without evidence.

UX Documentation

Lead

Define:

• who the document is for,
• which flow it describes,
• which terminology applies,
• which information must not be invented.

Expand

Ask for:

• a structure,
• the happy path,
• failure paths,
• dependencies,
• open questions.

Refine

Remove:

• non-existent system behaviours,
• duplicates,
• unconfirmed requirements,
• unnecessary explanations.

Elevate

Add:

• the rationale,
• user impact,
• the most important risk,
• decisions required from the team.

I will explore this topic more deeply in the next article: UX Documentation With AI: Prompts for Flows, Decision Logs, and Handoffs.

Writing for Stakeholders

Lead

Explain:

• who the recipient is,
• what they need,
• which decision they must make,
• what level of detail is appropriate.

Expand

Generate:

• a concise version,
• a detailed version,
• a problem–recommendation–risk structure,
• possible questions from the recipient.

Refine

Remove:

• jargon,
• repetition,
• decoration,
• unsupported promises.

Elevate

Add:

• your recommendation,
• the consequences of not making a decision,
• the priority,
• the required next action.

The goal is not for the stakeholder to think that AI produced an exceptionally elegant document.

The goal is to help them understand the situation and make a decision.

Collaboration After a Chaotic Meeting

Lead

Paste the notes and define:

• the subject of the meeting,
• the roles involved,
• the purpose of the summary,
• the agreed terminology.

Expand

Ask AI to separate:

• decisions,
• assumptions,
• questions,
• dependencies,
• responsibilities,
• next steps.

Refine

Verify:

• whether the decisions were genuinely made,
• whether AI assigned responsibility to someone incorrectly,
• whether the dates are correct,
• whether it added anything that was not said.

Elevate

Add:

• your understanding of the problem,
• your design recommendation,
• the point requiring confirmation,
• the consequence of leaving the issue unresolved.

“AI summarised the meeting” does not release anyone from the obligation to remember what actually happened during the meeting.

Portfolio Case Studies

In From a UX Prompt to a Portfolio Case Study, I described the complete process of building a case study around decisions rather than the chronology of screens.

The same workflow can help organise the material.

Lead

Provide the real project scope, your role, the constraints, and the available results.

Expand

Ask for several possible narrative structures.

Refine

Remove anything you cannot prove:

• invented metrics,
• research that did not happen,
• responsibilities that belonged to the team,
• business impact without evidence.

Elevate

Add:

• your point of view,
• key decisions,
• how your understanding of the problem changed,
• honest conclusions.

AI can help write a case study.

It cannot create professional experience you never had.

The Most Common AI Workflow Mistakes

1. Starting With Generation

The user writes:

Give me ten solutions.

But they have not established:

• what problem they are solving,
• which criteria will be used to evaluate the suggestions,
• which constraints exist.

The result is a broad list with little decision-making value.

Lead first.

Expand later.

2. Treating Possibilities as Recommendations

AI provides five edge cases.

That does not mean all five exist in the system.

AI compares two models.

That does not mean it knows the implementation costs.

AI presents “user needs”.

That does not mean it conducted research.

Expand produces material for evaluation.

Not a verdict.

3. Skipping the Refine Stage

The first AI response often looks good because it is:

• structured,
• complete,
• linguistically correct,
• decisive.

That is not evidence of accuracy.

Without Refine, you may publish:

• non-existent features,
• contradictory statuses,
• false assumptions,
• generic terminology,
• recommendations that do not fit the product.

The first response is a draft.

Even when it has headings.

4. Confusing Elevate With Decorating the Text

A more professional tone does not replace judgement.

Elevate does not mean:

• more adjectives,
• longer sentences,
• more uses of the word strategic,
• replacing “help” with “empower”.

Elevate means:

• explaining the choice,
• showing the consequences,
• adding real context,
• identifying the risk,
• taking responsibility for the recommendation.

5. Using AI Instead of Talking to the Team

When a question concerns:

• system behaviour,
• business rules,
• available data,
• technical dependencies,
• responsibility,
• security policy,

the answer must come from people who possess that knowledge.

AI can help formulate the question.

It should not impersonate a team member who knows the answer.

6. Failing to Control the Data

Do not paste the following without careful consideration:

• client data,
• transcripts containing personal information,
• documents covered by an NDA,
• internal requirements,
• financial information,
• health data,
• confidential communication.

A strong workflow is not only about response quality.

It is also about responsibility for the input.

A Template You Can Use at Work

You can apply the structure below to analysis, writing, documentation, or decision-making.

LEAD

Context:
What am I working on? Who uses the product? What stage has the project reached?

Intent:
What result do I need?

Constraints:
What should AI not assume, change, or add?

Perspective:
How do I currently understand the problem? What is my hypothesis?

EXPAND

Ask for:

• alternatives,
• perspectives,
• risks,
• scenarios,
• edge cases,
• comparison criteria,
• possible consequences.

REFINE

Verify and remove:

• unsupported assumptions,
• out-of-scope elements,
• repetition,
• incorrect terminology,
• unrealistic recommendations,
• information requiring confirmation.

Prioritise the remaining material.

ELEVATE

Add:

• your decision,
• the rationale,
• trade-offs,
• organisational context,
• impact on the user and system,
• unresolved questions,
• responsibility.

How to Determine Whether the Result Is Ready

Before using the material, answer seven questions:

1. Do all the facts come from a reliable source?
2. Has AI filled any gaps with its own assumptions?
3. Does the result address the actual objective?
4. Is the terminology consistent with the product?
5. Can I explain every recommendation?
6. Are the constraints and unresolved questions visible?
7. Does the final decision still belong to me or the appropriate person on the team?

If your answer to the final question is “AI selected the best option”, the workflow has just left the room without saying goodbye.

A Good Workflow Does Not Restrict Creativity

Structure may appear to be the opposite of creativity.

In practice, it does the opposite.

It allows you to:

• begin work faster,
• explore more possibilities,
• identify weak points earlier,
• separate ideas from decisions,
• preserve your own voice,
• build a repeatable standard of quality.

Without a process, every conversation with AI begins from zero.

The results depend on:

• your mood,
• the urgency,
• the accidental wording of the prompt,
• whether the first response looks sufficiently polished.

A workflow introduces control.

Not control over every generated sentence.

Control over the direction of the work.

AI Should Not Lead the Project

AI can:

• expand,
• organise,
• compare,
• edit,
• summarise,
• identify possible gaps.

The designer still needs to:

• define the problem,
• establish priorities,
• understand the context,
• verify the evidence,
• account for constraints,
• make or recommend the decision,
• explain the consequences,
• take responsibility.

The most important distinction is:

AI generates possibilities. The designer gives them meaning.

The Lead → Expand → Refine → Elevate model is not intended to produce more impressive answers.

It is intended to preserve authorship.

At the beginning, you establish the direction.

In the middle, you use AI’s speed and ability to organise information.

At the end, you restore context, judgement, and responsibility.

This is not a method that allows AI to design for you.

It is a method that allows you to use AI without ceasing to be a designer.

---

Do You Want to Build Your Own System for Working With AI?

The Designer’s AI Playbook develops the Lead → Expand → Refine → Elevate model and includes a library of 50 UX prompts, practical workflows, and templates supporting:

• design-decision analysis,
• research and information synthesis,
• UX documentation,
• user-flow descriptions,
• decision logs,
• communication with PMs, developers, and stakeholders,
• case-study writing,
• project-presentation preparation.

This is not a book about finding one perfect prompt.

It is a working system for designers who want to use AI to extend their own thinking without surrendering direction, authorship, or professional judgement.

Explore The Designer’s AI Playbook →