Open with the pair you brought
b1 ended with homework: generate the same screen twice, once from an adjective brief and once from a constraint brief, and keep both. Put them side by side now. In almost every case the adjective version is prettier at first glance and the constraint version is the one you could actually take into a review, because every claim in it is checkable. That gap is the whole session. Tonight we make the constraint brief a repeatable structure with six slots, then get honest about the one thing designers most want from these tools and cannot have: a reference image that guarantees brand compliance.
Anatomy of a constraint brief 14 min live
Six slots. Fill all six and the output is directable; skip one and the model fills it with the most likely thing, which is another way of saying the most generic thing. The test for each slot is the same: could a stranger check whether the output obeyed you? If the answer is no, that slot is still an adjective wearing a constraint's clothes. Four of the six map straight onto checks the critique lab measures, and two of them sit in the judgement half where no lab can help you. Both kinds belong in the brief.
LiveSlot 1 · Job - what decision does this screen serve2 min▶
Not "a summary screen". A screen exists so somebody can make a decision or complete a task. For Cadence's shared summary the job is: the person who missed the meeting decides whether they need to read the full transcript, and whether anything is theirs to do. Write it as a sentence with a subject and a verb.
One-line test: does the sentence name a person and a decision? "Show the meeting summary" fails. "Let a non-attendee decide in fifteen seconds whether anything here is theirs" passes, because you can hold a generated screen up against it and disagree with evidence.
LiveSlot 2 · One primary action2 min▶
Exactly one. For the shared summary it is Share summary. Everything else - copy, edit, export, view transcript, regenerate - is secondary and must be visually quieter. Two primary actions is the single most common structural failure in generated UI, because the model has seen a million screens where every button looks important.
One-line test: can you point at the loudest element and name it before the screen finishes loading? If two candidates tie, hierarchy is broken and the lab will say so.
LiveSlot 3 · Structure - scale, base size, unit, grid3 min▶
Four numbers, and they are arithmetic rather than taste. State the type scale as a ratio and a base, for example a 1.25 ratio from a 16px body, which gives the lab's ladder of 12 / 14 / 16 / 20 / 25 / 31 / 39. State the spacing unit, for example 8px, and say that every gap and padding is a multiple of it. State the grid: column count and gutter. Say all of this in the brief and the generator has no room to invent six unrelated font sizes.
Be honest with your team that the ratio and the unit are conventions, not standards. Nothing requires 1.25 or 8. What matters is that one ratio generates all your sizes, so the set stays small and related, and that one unit generates all your spacing, so the rhythm is visible.
One-line test: could someone recompute every size and gap in the output from your brief alone? If not, add the missing number.
LiveSlot 4 · Colour - one accent, neutrals, and a contrast floor3 min▶
One accent hue with a hex value, a neutral ramp, and a stated floor: all text passes WCAG 2.2 SC 1.4.3 at 4.5:1 for normal text and 3:1 for large, where large means 18pt or 24px, or 14pt and 18.66px when bold. Add SC 1.4.11 if the screen has controls: 3:1 on the visual boundaries of components and meaningful graphics, which is why a pale grey button outline is a defect rather than a style choice.
The floor is the part people forget, and it is the part you can enforce without a taste conversation. It is also the reason "add colour" is so expensive later: a second and third accent arrive without a contrast budget and the failures follow them in.
One-line test: is there a number in this slot? A hue name is a mood. A hex plus a ratio is a constraint.
Self-studySlot 5 · References, each with a stated reason3 min read▶
A reference without a reason is a mood board with ambitions. Say what you are borrowing and, more importantly, what you are not. "This one for the density of the metadata row, not the colour." "This one for the restraint in the header, not the illustration style." Three references with reasons beat twelve without, because a generator averages what you give it, and twelve unlabelled references average into the same generic middle you were trying to escape.
One-line test: for each reference, can you finish the sentence "for the ___, not the ___"? If not, drop it. Part 2 covers what the reference parameters can actually transfer, which is less than most people assume.
Self-studySlot 6 · Exclusions - what must NOT appear3 min read▶
The slot nobody writes and everybody needs. Generated output has a recognisable signature, and exclusions are how you refuse it up front. For Cadence's shared summary: no gradient background, no centred hero layout, no decorative iconography that does not encode meaning, no stock photography of people in meetings, no more than one accent hue, no drop shadows deeper than the elevation token, no faux browser chrome, no lorem text in any state.
Exclusions do two useful things. They cut the average out of the output, and they turn a vague review comment into a brief violation you can point at without discussing taste.
One-line test: would a reviewer be able to say "that is on the exclusions list" and be immediately right? If the item is arguable, make it specific or drop it.
Reference discipline, and the promise nobody makes 11 min live
Every designer asks the same question in week one: can I feed it our brand and get brand-compliant output. The honest answer is in the vendors' own documentation, and it is no. Reference parameters buy family resemblance. They do not buy compliance. Midjourney's docs put it plainly: references are used as inspiration and not to copy exactly. Reading those limits properly is what stops you burning a week and a credit budget on a promise that was never made.
| Parameter | What it actually does | Range and version | The documented limit |
|---|---|---|---|
--sref with --sw | Style reference. Transfers the overall style of the referenced image or code | --sw 0-1000, default 100. --sv selects the style-reference model | The docs are explicit that it "doesn't copy objects or people, just the overall style" |
--cref with --cw | Character reference, for keeping a character consistent | --cw 0-100. V6 only | Not available on current versions - if your team is on V7 or later this is not a route |
--oref with --ow | Omni Reference, V7 and later. Takes exactly ONE image | --ow 1-1000, and the guidance is to stay below 400 | Costs 2x GPU time, so it eats a metered budget twice as fast |
--p personalization | Applies a profile learned from your own ratings | Profiles must be unlocked by rating images first | Moodboard profiles cannot be combined with --sv or --sw |
| Style codes | Reproducible styles drawn from Midjourney's own library | Used in place of a --sref image | You cannot create a style code from an uploaded image, so your brand cannot become one |
LiveWhy "make it look exactly like our brand" cannot be answered here4 min▶
Put the four documented limits in a row and the conclusion writes itself. A style reference moves style and not objects, which rules out logos, wordmarks and product furniture. Character reference is a V6-only feature, so consistency of a specific figure is not a current-version route. Omni Reference takes a single image and doubles GPU cost, which makes it a precision instrument rather than a bulk one. And style codes come from Midjourney's library, which means the one thing you actually want - a code that encodes your brand from your own files - does not exist.
So the honest framing for a designer: reference parameters get you into the right family. Compliance comes from somewhere else. Somewhere else is two things, both later in this track. b6 puts your decisions into tokens and components, where the compliant option is the easy option and a machine can enforce it. b5 turns the five checks into a review script, which is how anything generated gets held to the standard before it ships. Direction gets you close; systems and critique get you correct.
Self-studyHow to write a reference brief that survives this3 min read▶
Given the limits, use references for what they are good at and put the rest in words and numbers:
- References carry mood, texture, density and light. Let them. Say which of those you want from each image.
- Words and numbers carry structure and colour. Your type ratio, base size, spacing unit, accent hex and contrast floor belong in the prompt as values, not in an image the model will average.
- Anything with a brand mark on it stays out of the generated layer entirely. Composite logos in afterwards. This is also a guardrail the leader track writes down in a6.
- Watch the weights. A high
--swpushes style hard and can flatten the structure you specified. Start at the default and move one variable at a time so you can attribute the change. - Log what you used. Model, version, parameters, reference images and their licences. b7 makes the provenance case in full, but the habit starts the first time you generate anything you might ship.
All of the above was verified from official documentation on 2026-08-08. This part of the market changes monthly, so check the current parameter pages before you teach a number to anyone else.
Direct, do not polish 8 min live
When the output is wrong, there are two moves. You can nudge the pixels - resize, recolour, move things until it stops annoying you. Or you can name the constraint you failed to write, add it, and regenerate. The first move feels like progress and is the most expensive habit in design, because you are paying craft time to improve a wrong idea. The second is cheap, repeatable, and leaves you with a better brief for next time. Here it is on Cadence's shared-summary screen, three rounds.
"Design a clean, modern shared-summary screen for Cadence, our AI meeting-notes app. Make it feel premium, trustworthy and a bit friendly."
A centred hero, a gradient band, a large heading in a default-feeling typeface, six font sizes with no relationship between them, three accent colours because nothing decided, and a pleasant illustration of a laptop. Nothing in it is wrong, because nothing in it was specified. On the lab's five checks this class of draft scores 17 out of 100.
"Shared-summary screen for Cadence. Job: a person who missed the meeting decides in fifteen seconds whether anything here is theirs to do. Exactly one primary action: Share summary. Copy, edit, export and view transcript are secondary and must be visibly quieter. The meeting title is the loudest element on the screen; the AI-generated label must be visible but must not compete with it. Left-aligned, no centred hero."
"Keep round 2. Type on a 1.25 scale from a 16px body, so sizes come only from 12 / 14 / 16 / 20 / 25 / 31 / 39. Every gap and padding is a multiple of 8px. One accent hue, plum #6B2D5C, plus a neutral ramp; no second or third accent. All text meets WCAG 2.2 SC 1.4.3, 4.5:1 normal and 3:1 large; control boundaries meet SC 1.4.11 at 3:1. Exclusions: no gradient, no decorative icons, no stock photography, no lorem text in any state."
LiveThe follow-up that decides whether you are art-directing3 min▶
Round 3 comes back close but not right. The share button is still competing with the title. Here are the two possible next messages, and only one of them is design work.
"Make the button a bit smaller and move it down. Maybe lighten the colour slightly. Also nudge the title up and add some spacing above it, and let's try the button in a softer shade."
"The brief said the meeting title is the loudest element. It is not - the primary button currently reads first. Add to the brief: the title is at least 1.5x the size of any button label and carries the highest contrast on the screen; the primary button sits below the summary body, never above it. Regenerate from the updated brief."
The first message produces a slightly less annoying version of a screen with a hierarchy problem, and teaches you nothing you can reuse. The second fixes the class of error, improves the brief permanently, and can be checked by someone who was not in the room. Same effort, entirely different asset.
Brief your way up the ladder ★ 15 min · everyone runs it
Same lab as b1, different job. Last session you toggled levers and watched the score. Tonight you work backwards: for every rung you climb, write the brief clause that would have produced it without any toggling at all. The canon numbers are 17 → 32 → 46 → 66 → 86 → 100, and each step is one clause you failed to write in round 1. By the end you should have a six-slot brief on paper that, if you had sent it first, would have skipped the whole climb.
Start with every lever off: 17. Read the five failure rows and write, in one sentence each, the missing brief clause. That sheet is your slot inventory.
Turn on Hierarchy only: 32. Your clause is slot 2. Write it with a number in it, for example "the title is at least 1.5x any button label and carries the highest contrast".
Add Type scale then Spacing: 46 then 66. Both clauses are slot 3, and both are arithmetic. Name the ratio, the base size and the unit rather than describing a feeling.
Add Contrast: 86. This is slot 4's floor, and it is the one clause you can cite by standard number - WCAG 2.2 SC 1.4.3, with SC 1.4.11 for control boundaries.
Add One accent: 100. Slot 4 again, plus slot 6. Then press Add colour and watch the finished screen lose 19 points - that is what a missing exclusions list costs.
Keep the brief, not the render. The most valuable artifact from a week of generating is almost never a file. It is the six-slot brief that finally produced usable work, because that brief transfers to the next screen, the next feature and the next person who joins the team. Renders expire; direction compounds. By b10 that brief is the front page of your handoff.
Try it yourself - this week ◐ 40-50 min total
- Take a real brief you have actually sent and rewrite it into the six slots. Mark any slot where you cannot pass the one-line test, and fix that slot before you send anything.
- Run the same screen from both briefs, the original and the six-slot version, in whatever tool you use. Keep the pair. b3 opens with it, and by b5 you will score both.
- Write the exclusions list for Cadence's shared summary. Aim for eight specific items a reviewer could point at without arguing, and keep it - it becomes part of your capstone in b10.
- Do one full round where you are forbidden from touching the output. You may only change the brief and regenerate. It will be uncomfortable for about ten minutes and then it will be how you work.
Official sources covered
Every parameter, range and limit on this page was taken from official documentation on 2026-08-08. Version numbers, credit costs and beta status in this space change monthly, so re-check the vendor's own page before you teach a number to your team. This page covers:
--sref with --sw 0-1000 default 100, --cref with --cw 0-100 V6 only, --oref V7+ single image with --ow 1-1000 at 2x GPU timePart 2 · full table, with the documented limits beside each--p profiles unlocked by rating, moodboards not combinable with --sv or --swNamed here · imagery workflow in full in b7Three questions before you go 🎯 ◐ 90 seconds
1 · According to Midjourney's own documentation, what does --sref do and not do?
Style references move style, not content. The docs state it "doesn't copy objects or people, just the overall style", and separately that you cannot create a style code from an uploaded image. Family resemblance yes, brand compliance no - that comes from tokens in b6 and your review in b5.
2 · Your brief names the job, the primary action, the type scale, the spacing unit and one accent. Which slot is still missing?
Six slots: job, one primary action, structure, colour, references with stated reasons, and exclusions. Exclusions is the slot nobody writes, and it is how you refuse the generic signature up front instead of arguing about it in review.
3 · Round 3 comes back with the share button competing with the meeting title. What is the art-direction move?
Fixing the pixels improves one file once and teaches you nothing reusable. Fixing the constraint repairs the whole class of error, improves the brief permanently, and can be checked by someone who was not in the room. Direct, do not polish.
Designer session 2 cheat sheet · pin this
--sref + --sw 0-1000 style only · --cref + --cw 0-100 is V6 only · --oref V7+ one image, --ow 1-1000 under 400, 2x GPU time.