Divergent exploration has to be forced 8 min live
Design has two motions. Divergent: widen the field, collect options you would not have thought of, stay deliberately unresolved. Convergent: narrow, decide, commit. A generator is convergent by construction - it returns what is most likely given your words, which means it hands you the middle of the category on the first try and every try after that. That is genuinely useful when you want the middle. It is fatal when you are doing research, because you get forty renderings of one idea and mistake volume for exploration.
LiveFour techniques that actually widen the field5 min▶
Each of these is a way of taking the decision away from the model's default. Use them one at a time so you can read the result.
- Vary one axis at a time. Pick the axis by name - structure, tone, era, medium, density - and hold everything else still. Five outputs along one axis teach you more than fifty along none, because the difference between them is legible.
- Brief against the obvious. Write down the cliche of your category first, then forbid it in the brief. "No hero gradient. No centred stack. No abstract swoosh. The AI badge is not a sparkle." Naming the cliche is what lets you check whether you escaped it.
- Pull references from outside the category. If every reference is a SaaS dashboard you will converge on a SaaS dashboard. Rail timetables, museum labels, boarding passes and print editorial all solve density and hierarchy problems that product UI keeps re-solving badly.
- Generate deliberately bad options. Ask for the worst defensible version, the most cluttered, the most austere. Boundaries are information: they show you where the acceptable range actually ends, and they are cheap now.
The axes are worth memorising because they double as review language. "This is a tone change, not a structure change" ends a lot of circular meetings.
Self-studyDivergent time is not the same as generation time3 min read▶
Cheap generation compresses the making, not the deciding. If you generate for twenty minutes and decide in two, you have not explored - you have shopped. A workable split for a session like tonight's:
- Diverge on paper or in words first. List the axes and the values before you open a tool. This is the part the tool cannot do, because it does not know what you are trying to avoid.
- Generate along the named axis. Small batches. Five per value, not fifty.
- Converge with a written reason. "This one, because it makes the correction step visible" - a sentence you can defend, not a shortlist you feel good about.
Keep the rejects. They are the only evidence that a choice was a choice, and Part 3 turns them into a working defence.
A moodboard is an argument, not a mood 7 min live
The unannotated board is the most common wasted artifact in design. Twelve beautiful references, no reasons, and every viewer takes away something different - which is why the feedback comes back as "love the vibe" and the work comes back wrong. An annotated board states a position: for each reference, what it is here for, what it is explicitly not here for, and the constraint it turns into. That third column is the whole point, because it is the only part that survives into the brief.
"Moodboard for Cadence shared summary - 12 refs. Vibe: calm, premium, intelligent. Let me know what you think!"
"Cadence shared summary - 5 refs, each with a reason and a rejection.
1. Swiss rail timetable · FOR its type hierarchy at high density · NOT for its cold neutrality · BECOMES: three type sizes, no more.
2. Airline boarding pass · FOR making status scannable in one glance · NOT for its cramped spacing · BECOMES: one state chip per row.
3. Museum wall label · FOR quiet attribution next to the thing itself · NOT for its serif voice · BECOMES: AI credit line, small, grey, always present.
4. Newspaper correction column · FOR treating a fix as normal, not shameful · NOT for its density · BECOMES: 'edited by you' is a label, not a warning.
5. Rejected: dark-mode analytics dashboard · charts imply precision this summary does not have.
Forbidden in this direction: hero gradient, centred stack, sparkle icon for AI."
LiveFirefly Boards, honestly4 min▶
Adobe's Firefly Boards is the mainstream AI-native moodboard surface, so it is worth knowing what is true about it:
- It is generally available worldwide, and anyone with an Adobe account can access it. You do not need a Creative Cloud plan to open a board.
- Unlimited boards are only on plans that include premium features. Other plans are limited to a small number of boards, so a free account is a place to try the surface, not a place to keep a year of research.
- Partner models inside Boards require a premium-feature subscription. This matters more than it sounds: partner models are how you get outside Adobe's own look, which is exactly what divergent exploration needs.
And the part no vendor will tell you: the tool matters far less than the annotation discipline. A Figma frame, a Firefly board, a shared folder and a physical wall all work if every reference carries a reason and a rejection. None of them work without that. Pick the surface your team will actually open, then spend your energy on the columns.
Self-studyHow to run the board past a stakeholder3 min read▶
Present the constraint column, not the images. Read three rows out loud in the form "we are taking X for Y, and deliberately not taking Z". Two things happen: the conversation moves off taste, and disagreement arrives early and specific, which is the cheapest place for it to arrive.
Then show the rejected pile briefly. Stakeholders who see what you turned down stop asking whether you considered it, and that single move removes most of the "have you thought about..." loop later in the project.
Why everything is starting to look the same 6 min live
The argument here is mechanical, not statistical, and it is worth being careful about that: a generator returns what is most likely given the words it was given. If you and your competitors brief the same tool the same lazy way, the most likely output for all of you is the same output. No study is needed for that claim, because it is a description of how the tool works. What you can also do is look, and note the signature that generated drafts share.
LiveThe generated signature - what to look for3 min▶
These are observable properties of output, not claims about the industry. Hold a generated draft next to this list:
- Centred everything. Centre is the safest composition, so it is the likeliest one. Asymmetry requires a decision.
- One big gradient. Usually behind a hero, usually doing no work. It reads as "something is happening here" without anything happening.
- Decorative iconography that means nothing. A sparkle for AI, an abstract swoosh, three icons that could be swapped without changing the meaning.
- Three accents because nobody decided. The colour count is the tell. Undecided colour is the most legible symptom of an undecided brief - and it is the one thing the critique lab prices directly.
None of these are crimes on their own. Together, on a first draft, they mean the model chose and you did not.
LiveThree defences a working designer can actually run4 min▶
- Arrive with a position the tool does not have. The board's constraint column is that position. It is specific to your product, your user and your brand, so it is not in the model's average - which makes it the only reliable source of difference.
- Keep a rejected-options file. One folder or one page per project: what you generated, what you turned down, and the one-line reason. It is proof that exploration happened, it stops you re-walking the same ground in month three, and it is the fastest way to answer a late "why not like this?".
- Audit against the category, not against your last screen. Once a quarter, put your screen beside five competitors and one thing from outside the category. Self-comparison hides convergence, because you drift together. This is the only check that catches it.
Notice that none of the three is "use the tool less". The tool raises your floor. The defences protect your ceiling.
The fastest version of the category audit takes fifteen minutes. Screenshot your screen and five competitors at the same width, put them in one row, and ask a colleague to name which one is yours with the logos cropped out. If they cannot, you have a direction problem and no amount of polish will fix it. If they can, ask them what gave it away - that answer is your actual brand asset, and it belongs in the next brief.
Explore one axis, then annotate five references ★ 12 min · everyone runs it
We do both halves on Cadence's shared-summary screen. First force divergence along structure - the axis from the diagram - then build the board that turns the winner into constraints. Use whatever tool you have open; the discipline is the deliverable, not the surface.
Write the cliche down first: "hero gradient, centred stack, sparkle for AI, three accents". This is now forbidden text you can paste into every prompt tonight.
Name the axis - structure - and generate the five values from the diagram: timeline, decisions-first, one card per speaker, single paragraph, and the deliberately bad raw chat log. Tone, era, medium and density stay locked.
Pick one, and write the reason in a sentence that could be wrong. "Decisions-first, because the person who missed the meeting wants the outcome, not the transcript."
Build the five-row board: reference, what it is for, what it is not for, the constraint it becomes. At least two references must come from outside product UI.
Start the rejected file. Four structures and their one-line reasons for rejection, saved next to the board. That file is your evidence for the rest of the track.
One more thing before you go and be distinctive: exploration is upstream of craft, not a substitute for it. The lab below is the same ladder from b1 - 17 → 32 → 46 → 66 → 86 → 100. A screen with a genuinely original concept still has to clear the same measurable floor, and turning on Add colour at the top still costs 19 points. Distinctive and broken is still broken.
Try it yourself - this week ◐ 30-40 min total
- Take one screen you are working on, name a single axis, and generate five variations along it only. Write one sentence on what the axis taught you that the first output did not.
- Build a five-reference annotated board with all four columns filled. At least two references from outside your category, and at least one reference you are rejecting on the record.
- Write your category's three cliches down and forbid them in writing in your next brief. Then check the output against your own list - honestly.
- Start the rejected-options file for Cadence's shared summary. Four rejects, four one-line reasons. Bring it to b4; you will draft against it.
Official sources covered
Tonight leans on the vendors' own documented behaviour for the moodboard surface, and on mechanism rather than statistics for the homogenisation argument. Where a claim would need a study we have not read, we say so instead of inventing a number. Tool access rules in this space change often - check today's page before you quote a limit to a client.
Three questions before you go 🎯 ◐ 90 seconds
1 · You generate forty options from one prompt. Why is that usually not exploration?
The generator is convergent by construction. Without a named axis you get variation in surface and none in idea. Five outputs across one deliberately varied axis teach you more than fifty from a single prompt, because the difference between them is legible.
2 · Which statement about Adobe Firefly Boards is accurate?
Boards is generally available worldwide and open to anyone with an Adobe account. Unlimited boards come with plans that include premium features; other plans are capped at a small number. Partner models inside Boards also need a premium-feature subscription - which matters because partner models are how you reach outside one vendor's default look.
3 · What makes a moodboard row usable in a brief?
Only the constraint column travels into the brief. "This for its type hierarchy, not its colour" is checkable; "calm and premium" is not. A row that produces no constraint is decoration, and a board built to please every taste in the room states no position at all.