SCROLL

AI PRACTICE · 2026

My AI Design Workflow

AI organizes evidence, compares directions, prototypes flows, and inspects the build. I define the problem and decide what moves forward.

My Workflow Stack

HOW I WORK WITH AI

`AI can accelerate the work. It cannot own the reasoning.`

I use AI to research, organize, generate, prototype, and inspect. It helps me move faster and explore more possibilities, but I remain responsible for the evidence, direction, quality, and final decision.

HOW I WORK WITH AI

`AI can accelerate the work. It cannot own the reasoning.`

GROUND IN EVIDENCE

Start with material you can trust.

AI organizes the material

Compares sources and surfaces gaps.

I verify the evidence

Confirm what is real, assumed, or missing.

GROUND IN EVIDENCE

Start with material you can trust.

AI organizes the material

Compares sources and surfaces gaps.

I verify the evidence

Confirm what is real, assumed, or missing.

SET THE DIRECTION

Explore widely, then choose deliberately.

Generate alternatives

AI reframes the problem and compares approaches.

Define & choose

I set the boundaries, weigh trade-offs, and choose the direction.

SET THE DIRECTION

Explore widely, then choose deliberately.

Generate alternatives

AI reframes the problem and compares approaches.

Define & choose

I set the boundaries, weigh trade-offs, and choose the direction.

APPLY THE QUALITY GATE

Generated does not mean finished.

Build & inspect

AI creates drafts and flags gaps or inconsistencies.

Approve, revise, or stop

I decide whether the work is ready to move forward.

APPLY THE QUALITY GATE

Generated does not mean finished.

Build & inspect

AI creates drafts and flags gaps or inconsistencies.

Approve, revise, or stop

I decide whether the work is ready to move forward.

THE END-TO-END PROCESS

`Workflow at a Glance`

Eight connected steps move the work from early context and evidence to a defined problem, a testable experience, a validated system, and reusable learning. Each step carries its decisions and outputs into the next.

THE END-TO-END PROCESS

`Workflow at a Glance`

Drag

THE WORKFLOW IN PRACTICE

`How Each Step Works`

Each step shows the same working relationship: how I direct AI, what it returns, and the judgment I apply before the work moves forward. Explore each action to see its prompts, deliverables, and decision gates in more detail.

THE WORKFLOW IN PRACTICE

`How Each Step Works`

Action 01

Establish the requirements context

Settle who, in what situation, is trying to do what. This stage produces no solutions, it only separates facts, preferences, assumptions and non-goals.

How I direct AI

Supply the known material, and draw the line between scope and non-goals
Make it separate what is fact, what is someone’s opinion, and what is still guesswork
Land every item on: for whom, when, what they want to do, and what counts as success

Supply the known material, and draw the line between scope and non-goals
Make it separate what is fact, what is someone’s opinion, and what is still guesswork
Land every item on: for whom, when, what they want to do, and what counts as success

Supply the known material, and draw the line between scope and non-goals
Make it separate what is fact, what is someone’s opinion, and what is still guesswork
Land every item on: for whom, when, what they want to do, and what counts as success

What AI returns

A structured first draft, plus the gaps it filled in without marking them: requirements no one asked for, statements no longer traceable to a person, guesses written as facts.

My judgment

Classify by who said it and where, not by how certain it sounds
Reject any requirement the AI invented without confirmation
Decide which unanswered questions block the next step

Classify by who said it and where, not by how certain it sounds
Reject any requirement the AI invented without confirmation
Decide which unanswered questions block the next step

Classify by who said it and where, not by how certain it sounds
Reject any requirement the AI invented without confirmation
Decide which unanswered questions block the next step

Deliverable 01

Requirements Context Brief

Four separate lists : facts, preferences, assumptions, non-goals. Every line carries who said it and where. Nothing is merged into one requirements section.

Criteria 01

Someone who missed the early discussions can say who the user is

They can point to a line that is someone's preference, not a confirmed fact

They can say what this round will not address

Action 02

Plan and organize the evidence

Turn “what is uncertain” into “how it should be confirmed.” Choose the question first, then the method.

How I direct AI

Decide which questions this round has to answer, then choose the method
Fence off which sources may be used, and over what time window
Require every conclusion to draw on at least two different kinds of source

Supply the known material, and draw the line between scope and non-goals
Make it separate what is fact, what is someone’s opinion, and what is still guesswork
Land every item on: for whom, when, what they want to do, and what counts as success

Decide which questions this round has to answer, then choose the method
Fence off which sources may be used, and over what time window
Require every conclusion to draw on at least two different kinds of source

What AI returns

A classification of the material, the patterns shared across users, and the counter-examples that contradict them — plus the parts it cannot judge true or false.

My judgment

Decide which questions are worth pursuing further
Count only what actually happened, never “it probably could”
Check that each pattern is genuinely backed by raw records

Decide which questions are worth pursuing further
Count only what actually happened, never “it probably could”
Check that each pattern is genuinely backed by raw records

Decide which questions are worth pursuing further
Count only what actually happened, never “it probably could”
Check that each pattern is genuinely backed by raw records

Deliverable 02

“Evidence Plan”

Lists the questions this round has to answer, pairs each with a concrete method, and states what result counts as enough.

“Traceable Evidence Map”

Every conclusion links back to its source. Sources come in three states, and each one looks different: real records, temporary stand-ins, still missing.

Criteria 02

Every significant conclusion can be traced back to a raw record

The counter-examples are still in the document, not buried under the conclusion

Explanations without support are labeled assumptions, not conclusions

Action 03

Define the problem and the goal

Find, in a mass of information, the problem that genuinely deserves design attention. A visible symptom is not the cause.

How I direct AI

Ask it to reframe the problem from several different angles
Compare how user, situation, obstacle and consequence relate
Flag the parts where the evidence is thin or the claim overreaches

Ask it to reframe the problem from several different angles
Compare how user, situation, obstacle and consequence relate
Flag the parts where the evidence is thin or the claim overreaches

Ask it to reframe the problem from several different angles
Compare how user, situation, obstacle and consequence relate
Flag the parts where the evidence is thin or the claim overreaches

What AI returns

Several candidate problem framings, each with its causal assumptions and design opportunities, plus the places where evidence is thin.

My judgment

Reject any problem that is only a phenomenon, with no consequence for the user
Never dress a predetermined solution up as a problem
Choose the direction that has evidence behind it and still allows more than one solution

Reject any problem that is only a phenomenon, with no consequence for the user
Never dress a predetermined solution up as a problem
Choose the direction that has evidence behind it and still allows more than one solution

Reject any problem that is only a phenomenon, with no consequence for the user
Never dress a predetermined solution up as a problem
Choose the direction that has evidence behind it and still allows more than one solution

Deliverable 03

"Problem Frame"

The chosen problem, its causal assumptions, the metric it affects, and the problems ruled out — every exclusion carries its reason.

Criteria 03

Someone who reads only this page can say what happens if the problem goes unsolved

They can say which metric it affects

They can say which problems were ruled out, and why

Action 04

Decide the experience strategy

Decide what change the experience should bring before discussing what the interface looks like.

How I direct AI

Set the dimensions for comparison
Ask for genuinely different experience mechanisms
Analyze the gains, risks, dependencies and trade-offs one by one

Set the dimensions for comparison
Ask for genuinely different experience mechanisms
Analyze the gains, risks, dependencies and trade-offs one by one

Set the dimensions for comparison
Ask for genuinely different experience mechanisms
Analyze the gains, risks, dependencies and trade-offs one by one

What AI returns

Candidate strategies compared by mechanism, with a risk list and the key assumptions that still need testing.

My judgment

Reject any option that changes appearance without changing understanding or behavior
Rank by fit to the problem, feasibility and impact
Pick a direction, and write down what it will not do

Reject any option that changes appearance without changing understanding or behavior
Rank by fit to the problem, feasibility and impact
Pick a direction, and write down what it will not do

Reject any option that changes appearance without changing understanding or behavior
Rank by fit to the problem, feasibility and impact
Pick a direction, and write down what it will not do

Deliverable 04

“Experience Strategy Statement”

States what change is expected in the user, the mechanism that delivers it, and what this direction explicitly will not do.

“Option Comparison Record”

Keeps the candidate directions, the shared comparison dimensions, and for each direction its gains, risks, dependencies, key assumptions and the reason it was rejected.

Criteria 04

Someone who missed the strategy discussions can say what change this experience is meant to produce

They can explain how the chosen mechanism answers the problem, not just describe what the interface looks like

They can say which directions were dropped, and why

Action 05

Build the structure and prototype

Turn the experience strategy into a task path that can actually be operated. The successful run is only part of it.

How I direct AI

Lock the core tasks and business rules so they stay unchanged
Ask for candidate information structures and flows
Check for omissions, break points and missing states

Lock the core tasks and business rules so they stay unchanged
Ask for candidate information structures and flows
Check for omissions, break points and missing states

Lock the core tasks and business rules so they stay unchanged
Ask for candidate information structures and flows
Check for omissions, break points and missing states

What AI returns

Information architecture and user flows, a page list and a state coverage report, plus a prototype that can be operated.

My judgment

Reorder the content hierarchy, and cut the redundant steps
Fill in the error states, empty states and recovery paths
Decide how much fidelity the prototype needs

Reorder the content hierarchy, and cut the redundant steps
Fill in the error states, empty states and recovery paths
Decide how much fidelity the prototype needs

Reorder the content hierarchy, and cut the redundant steps
Fill in the error states, empty states and recovery paths
Decide how much fidelity the prototype needs

Deliverable 05

“Information Architecture and Key Flows”

Shows the content hierarchy, the core task paths and the key decision points, and explains how the chosen experience strategy is carried into the structure.

“State Documentation”

Records the empty states, error states and recovery paths of the key pages, so the design is not limited to the ideal run.

“Testable Prototype”

Makes the core task operable. Fidelity goes only as far as the current question requires, without locking in the full visual design early.

Criteria 05

The core task runs from entry point to result, with no break in the main path

Empty states, error states and recovery paths are all explicitly recorded

Prototype fidelity is enough to test the current question, without moving into full visual design too early

Action 06

Establish the interface system and complete the build

This stage runs in the opposite order from most: the standards go up first, then AI produces inside them. Any page generated before the standards are fully bound is not accepted.

How I direct AI

Specify the content, functionality and experience principles that must not change
Extract design rules back out of reference images and existing pages, build them into variables, and bind them onto the existing pages
Only once the standards are stable, allow AI to generate new pages from the variables and components

Specify the content, functionality and experience principles that must not change
Extract design rules back out of reference images and existing pages, build them into variables, and bind them onto the existing pages
Only once the standards are stable, allow AI to generate new pages from the variables and components

Specify the content, functionality and experience principles that must not change
Extract design rules back out of reference images and existing pages, build them into variables, and bind them onto the existing pages
Only once the standards are stable, allow AI to generate new pages from the variables and components

What AI returns

Visual candidates and design rules, first-draft components and key pages, plus runnable code generated against the same set of standards.

My judgment

Reject visual options that differ only in decoration and carry no experience consequence
Choose the direction by reading order, accessibility and reusability
Check item by item that the code holds the design intent, and confirm the standards have landed in the code assets

Reject visual options that differ only in decoration and carry no experience consequence
Choose the direction by reading order, accessibility and reusability
Check item by item that the code holds the design intent, and confirm the standards have landed in the code assets

Reject visual options that differ only in decoration and carry no experience consequence
Choose the direction by reading order, accessibility and reusability
Check item by item that the code holds the design intent, and confirm the standards have landed in the code assets

Deliverable 06

“Visual Direction Board”

Answers what the interface should feel like, and why that direction was chosen. It sets the visual principles and the basis for choosing, so later design does not drift on personal taste.

“Interface System and Key Pages”

Turns the visual principles into variables, components, states and responsive rules. The key pages prove the rules can carry a real task, rather than staying an abstract design document.

“Runnable Front End”

Implements the key task path from the same variables and components, testing whether the design system holds up in real code, real interaction and different screens.

Criteria 06

The visual direction supports content hierarchy, reading order and accessibility, not just decorative change

Variables and components are bound back onto the existing pages, and every exception found is resolved or recorded

The runnable front end holds the design intent, and the standards have landed in the code assets

Action 07

Validate and iterate

Confirm separately whether the product was built correctly, and whether real users can finish the task. These are two different things.

How I direct AI

Separate the product check from the user validation first
Ask it to check the code, responsive behavior, keyboard operation and every state
Organize the problems that appear in real test records, without letting AI write in user feedback that never happened

Separate the product check from the user validation first
Ask it to check the code, responsive behavior, keyboard operation and every state
Organize the problems that appear in real test records, without letting AI write in user feedback that never happened

Separate the product check from the user validation first
Ask it to check the code, responsive behavior, keyboard operation and every state
Organize the problems that appear in real test records, without letting AI write in user feedback that never happened

What AI returns

Technical problems and experience problems listed separately, each with its severity, likely source and suggested fix.

My judgment

Never treat an automated check as user evidence
Explain and prioritize problems from real task performance
Decide whether to change the design, the task flow or the implementation

Never treat an automated check as user evidence
Explain and prioritize problems from real task performance
Decide whether to change the design, the task flow or the implementation

Never treat an automated check as user evidence
Explain and prioritize problems from real task performance
Decide whether to change the design, the task flow or the implementation

Deliverable 07

“Validation and Iteration Record”

Collects the problems found in the product check and the user validation. Each one keeps its original evidence, the task it affects, its severity, the fix decided and the retest result, with automated findings, real user evidence and still-unverified items kept clearly apart.

Criteria 07

Technical problems and experience problems are recorded separately, neither standing in for the other

Every significant judgment traces back to a real test record, not to AI inference or an automated check

Problems blocking the key task are fixed and retested; whatever is unresolved stays marked as unverified

Action 08

Deliver and consolidate

Deliver more than the current result: leave behind experience the next round can safely reuse.

How I direct AI

Check the delivery for completeness, and for the differences between versions
Draft the handover notes and the maintenance notes from the existing material
Pull out the rules, templates and working methods that might be reusable

Check the delivery for completeness, and for the differences between versions
Draft the handover notes and the maintenance notes from the existing material
Pull out the rules, templates and working methods that might be reusable

Check the delivery for completeness, and for the differences between versions
Draft the handover notes and the maintenance notes from the existing material
Pull out the rules, templates and working methods that might be reusable

What AI returns

A delivery checklist and change log, maintenance notes, and the candidates that could become reusable rules.

My judgment

Confirm who owns each delivered item and where its use stops
Separate what actually happened in the project from the reusable method templates
Refuse to dress one project’s experience up as a general conclusion

Confirm who owns each delivered item and where its use stops
Separate what actually happened in the project from the reusable method templates
Refuse to dress one project’s experience up as a general conclusion

Confirm who owns each delivered item and where its use stops
Separate what actually happened in the project from the reusable method templates
Refuse to dress one project’s experience up as a general conclusion

Deliverable 08

“Delivery Package”

Contains the final design, the design standards, source files, code, state documentation, the decision record, acceptance conditions and maintenance notes.