# Skill: Viral Carousel

Read this **before generating ANY slide** of a multi-image post: LinkedIn carousel, Instagram carousel, X thread cover set, "make me slides about X", "design a deck for this post". Also fetch `kubeez://skills/content-creation` (propose-before-generate + creative-director prompting). If the carousel is for a paid campaign or brand launch, also fetch `kubeez://skills/marketing-campaign`. The named-style gallery, palette presets, card treatments, type pairings and mark styles live in `kubeez://skills/carousel-style-previews` (the **style library**); fetch it whenever you use a named system.

## The bar (never changes, whatever the style)

A carousel slide is a **designed page with the text rendered into the image**, not "background + text on top". Every slide, in every style, clears four pillars:

1. **Topic-native visual world.** The subject of the post IS the world the slide lives in. Cars: carbon weave, headlight macro, stitching. Coffee: kraft bags, roaster labels, crema. SaaS: interface fragments, terminal, keycaps. Never a flat colour rectangle behind text.
2. **Typography as the artwork.** Text is composed, not laid out: scale contrast, a deliberate pairing, a clear first read. Name specific typefaces in prompts.
3. **A mark vocabulary, or its deliberate absence.** Hand-drawn marks, hairline rules, stamps, tape, UI glyphs, a signature device. Chosen once per deck. Never lucide-style sparkle / lightning / rocket icons.
4. **Texture or substrate.** Paper, grain, print, glass, fabric, screen. The surface is part of the design. Only a clean Swiss / UI style may run near-flat.

**Default lean: layered and scroll-stopping, not gallery-austere.** Slides compete in a feed. A hook slide should carry at least three layers (substrate, image or graphic, headline, mark). One word on a blank page and a zero-text photo read as art prints, not as posts. Restraint is a choice for luxury / premium-B2B topics, not the default.

Auto-fail looks: centered subject on a plain gradient, white serif on solid black, stock-photo-with-caption, neon glow, holographic "AI" imagery, circuit boards, lens flares.

## Step 0: Freshness check (before any copy)

A polished deck built on stale or invented facts is worse than no deck. Decide whether the topic needs current data; if it does and the user did not supply the specifics, **web-search first**.

- **Always search** (when the user gave no details): news, geopolitics, elections; markets, prices, earnings; sports results and transfers; product / model launches, versions, pricing, benchmarks; entertainment numbers, release dates, awards; disasters and recalls; "today in history" framing; anything with "this week / latest / new / breaking"; any number, date or quote the user did not spell out.
- **No search needed**: how-to, frameworks, opinion, craft, settled history, personal narrative the user provided in full, or topics where the user pasted the source material.
- **How**: at least two queries (one broad, one targeted), today's actual date in the query, wire services / primary sources first. Cite sources as markdown links in the approval step. Never invent a number or a quote: drop the slide or use sourced phrasing ("hundreds of strikes since February"). A plain dateline ("May 11, 2026") on time-sensitive slides is honest; an invented serial kicker is not.

Write a 3 to 5 line factual brief and get it approved with the slide narrative before composing the design.

## Step 1: Pick the design mode

| Mode | When | Go to |
|---|---|---|
| **Reference** | The user attached a finished deck / slide / ad and says "like this", "recreate", "same style for X", "rebrand this" | Step 2A |
| **Named style** | The user named one of the 9 library styles, or tapped one in the style picker | Step 2B |
| **Art director** | The user has no style preference: "you choose", "just make it", "surprise me", tapped **You choose for me** ("You pick the composition system that fits this topic best."), or declined the gallery | Step 2C |

**When the user started a deck and said nothing about style:** offer the choice once. In the Kubeez chat, emit the `<style_picker>` block (format in the style library; its "You choose for me" button routes to Art director). In other clients, list the 9 styles as name + one line + preview link and add "or I design a bespoke look for this topic". If the user's brief already signals they want you to decide (a complete brief, "go", "just do it", no interest in options), skip the offer and run Art director. Never show the gallery twice in one deck.

## Step 2A: Reference recreation

The reference IS the design system. Paraphrasing an attached layout into an abstract design block destroys it (observed 2026-07-08: the rewrite never mentioned the attached images, lost the layout, and invented a brand colour).

1. Skip Art director and the library. Derive the design block by DESCRIBING the reference precisely: its real palette, type behaviour, grid, marks, texture.
2. Every slide prompt opens with recreation language ("Match the attached reference slide's layout exactly: same grid, type hierarchy, mark vocabulary") and **manifests every input image**: `IMAGE 1 is the layout reference to match. IMAGE 2 is the official <brand> logo, use it exactly.` An attachment the prompt does not explain degrades into a vague style hint.
3. Brand colours are the brand's REAL hexes (brand kit, user assets, the brand's own site), never guessed from the brand's name. Copy is adapted to the brand's actual business, not the reference brand's claims.

## Step 2B: Named style

Fetch `kubeez://skills/carousel-style-previews` for the chosen style's full description, preview, palette presets, card treatments, type pairings and mark styles. Lock one palette, one card treatment, one type pairing and one mark vocabulary for the whole deck. Then use the style's row of the overlay table in Step 4 for every slide prompt. Match the style's preview; do not drift it toward another style.

## Step 2C: Art director (design a bespoke system for THIS topic)

The 9 library styles are examples, not a menu. In this mode you design a system that only makes sense for this post. **The core move: the deck looks like it was printed by the world it talks about.** Every industry already has its own graphic culture. Borrow its artefacts instead of a generic magazine look.

### 1. Find the source artefact

List three to five artefacts the topic's world actually prints, wears, displays or packages. Pick the one with the most graphic character and enough surface for text.

| Topic | Source artefacts to borrow from |
|---|---|
| Motorsport / cars | Livery stripes and sponsor-decal type, timing-screen numerals, tyre sidewall lettering, pit-board signs |
| Coffee / food / restaurants | Roaster bag labels on kraft, menu design, cupping score sheets, receipt printer tape, chalkboard specials |
| Finance / fintech | Annual-report tables, banknote guilloche and serial numbers, ticker tape, cheque layouts |
| SaaS / dev tools / AI | Changelog pages, terminal output, keycaps, release notes, error dialogs, API docs |
| Fitness / sport | Gym whiteboard workouts, jersey numbers, race bibs, chalk on rubber flooring, scoreboards |
| Skincare / beauty | Pharmacy labels, ingredient lists, frosted glass bottles, lab test strips |
| Real estate / architecture | Floor-plan linework, listing sheets, blueprint dimension marks, brass door plates |
| Music / nightlife | Gig posters, setlists taped to a stage, vinyl centre labels, wristbands, flyers |
| Travel / hospitality | Boarding passes, luggage tags, postcards, hotel key sleeves, passport stamps |
| Fashion / apparel | Care labels, hangtags, lookbook spreads, pattern-cutting paper, woven labels |
| Education / how-to | Exam papers, textbook diagrams, flash cards, lab notebooks |
| News / politics | Wire photos, front pages, briefing folders, ballot papers, court sketches |
| Health / medical | Prescription pads, lab reports, anatomy plates, appointment cards |
| Gaming / entertainment | Cartridge labels, HUD overlays, key art, cinema tickets, trading cards |

If the topic is not listed, ask "what does this industry hand to its customers, hang on its walls, or show on its screens?" and pick from that.

### 2. Draft two contrasting concepts, keep one

Write two one-line concepts built on different artefacts and different media (one image-led, one type- or graphic-led). Example for "7 mistakes new coffee shop owners make": *(a) Cupping Sheet: each slide is a scored cupping form, mistakes marked in red pen; (b) Bag Label: each slide is a roaster's kraft bag label, the mistake as the origin line.* Pick the one that best carries the post's voice, and note in one line why the other lost. Offer the runner-up to the user only if they seem undecided.

### 3. Fill the spec (every axis, concrete values)

| Axis | Write down |
|---|---|
| **Name** | A short name for this system ("Cupping Sheet"). Used in the pitch and in `log_carousel_deck`. |
| **Source artefact** | The real-world object the slides imitate, and how literally (full imitation vs borrowed details). |
| **Medium** | Photograph, illustration, 3D studio render, collage, type-only, UI surface, or a named mix. The most load-bearing axis. |
| **Layout grid** | Where the eye lands first, second, third. Margins. Centred, ranged left, or artefact-driven (form fields, label panels). |
| **Type** | Class + named families + the pairing logic ("Druk Wide for scores, IBM Plex Mono for form labels, handwritten red pen for notes"). Taken from the artefact where possible. |
| **Palette** | 2 to 4 hexes sampled from the artefact and the topic's world, not from a preset. Never pure black or pure white. |
| **Substrate + texture** | The surface: kraft paper, thermal paper, frosted glass, screen, fabric. Its grain, print artefacts, wear. |
| **Signature device** | The ONE recurring graphic element unique to this deck (a red-pen circle, a perforated tear line, a timing-screen digit). It appears on every slide and ties the deck together. |
| **Image treatment** | How photos or objects enter: printed onto the artefact, clipped to it, cut out on top, or none. |
| **Marks** | The deck's mark vocabulary (stamps, pen marks, rules, UI glyphs). One vocabulary. |
| **Continuity** | How one slide leads to the next: repeating header band, a running number, the device moving position. |

Optional reach: a **direction family** can sharpen the concept. Print ephemera (labels, tickets, receipts). Poster / colour-block (flat spot colours, oversized type, riso misregistration). Cinematic (film still, letterbox, subtitle-style type). Data-led (one chart or number as the hero, annotated). UI-native (app screens, notifications, search results, maps). Object studio (one symbolic object on a seamless, type around it). Hand-made (collage, marker, cut paper). Swiss / editorial (grid, rules, restraint). Brutalist type (huge raw type, hard edges, no ornament). Mix them only if the artefact demands it.

### 4. Run the four tests before pitching

1. **Swap test.** Replace the topic words with another industry's. If the design still works unchanged, it is a template, not a bespoke system. Push the artefact harder.
2. **Rename test.** If it is a library style (The Press, The Moleskine...) with new colours, you are defaulting. Use Named style honestly or design something new.
3. **Layer test.** The hook slide has at least three layers and one clear first read.
4. **Thumbnail test.** The headline reads at phone-thumbnail size. Artefact detail never beats legibility: the form fields, labels and fine print are background texture, the headline is huge.

### 5. Pitch it with the copy

In the approval step show: the system name, a two-sentence pitch ("Each slide is a cupping score sheet on cream card stock; mistakes are circled in red pen"), the palette hexes, the signature device, and the slide-by-slide copy. Get a yes before slide 1.

## Step 3: Deck structure

### Slide count (from topic depth, never a reflex 10)

| Topic depth | Slides |
|---|---|
| Hot take / one claim | 5 to 6 |
| Explainer / news / case story | 7 to 9 |
| Framework / step-by-step / multi-actor | 10 to 12 |
| Deep dive | 12 to 15, or split into Part 1 / Part 2 |

### Slide roles

The deck opens with a **Hook**, builds tension on slide 2, delivers value slides, lands one crystallised takeaway, and closes with a **CTA**. Slide 1 must work as a standalone post (it is the only slide most people see). Slide 2 must reward the swipe with the first real payoff, never a second intro.

### Archetypes

Each slide picks one archetype. Letters are internal shorthand; show users the NAME only. Fill `[SYS_HEADLINE]`, `[SYS_BODY]`, `[SYS_MARK]` from your system (the Step 4 overlay table for named styles, your Type / Marks axes for a bespoke system).

- **A. Hook** (slide 1): [SUBSTRATE] with the system's texture. [SYS_HEADLINE] rendering "[HEADLINE]" as the deck's loudest moment, max 8 words. One to three short support lines in [SYS_BODY]. One [SYS_MARK] or the signature device on the key word. Layout variants, pick one: centred stack; left-aligned with an oversized quote glyph; edge-bleed headline running off the frame; masthead (one full-width display line like a magazine nameplate, subtitle beneath, no kicker); one enormous word with a small caption; question hook with an oversized "?" as a graphic.
- **B. Editorial** (tension build): two-line stacked headline in [SYS_HEADLINE], two short paragraphs in [SYS_BODY], final line in [ACCENT] as a one-sentence takeaway. One mark.
- **C. Photo Overlay** (value slides): a topic-native photograph or object, specific to the post. How it enters depends on the system: full-bleed with a darkened text zone (Press, Photograph, cinematic), a grid cell beside the headline (Modernist, Swiss), a taped / clipped / printed-on print (Dossier, Moleskine, Billboard, most artefact systems), an attachment inside the UI card (Screenshot, UI-native). Headline + 2 to 4 short lines + one accent line.
- **D. Diagram** (framework / comparison / steps): headline, then EXACTLY [N] elements in the deck's card treatment, 2 short lines each, then a takeaway line. The fragile archetype: restate labels twice, keep N at 6 or less, expect a retry.
- **E. CTA Close** (last slide): one line of what-you-learned in [SYS_BODY], then the action in [SYS_HEADLINE] in [ACCENT], max five words (save, share, follow, comment a keyword, visit). Handle / URL only if the user asked for it.
- **F. Pull-quote** (optional, mid-deck): one large quote in the system's headline voice, near edge to edge. Attribution only when the source is real.
- **G. Big Number** (optional, proof): one stat rendered huge as the graphic, a one-line label, a one-line "so what". Sourced numbers only (Step 0).

**Rhythm rules.** No two adjacent slides share the same layout. Alternate dense and airy slides. The signature device appears on every slide but moves. Every slide carries the same substrate, palette and type, so the deck reads as one object when swiped.

## Step 4: Compose the DESIGN block (once, pasted verbatim into every slide)

Fill this once, get approval, then paste it unchanged at the top of every slide prompt. Paraphrasing between slides is the main cause of drift.

```
DESIGN SYSTEM (apply unchanged to every slide of this carousel):
• System: [NAME]. [Two-sentence description: the source artefact or named style, and how literally it is imitated.]
• Medium: [photograph / illustration / 3D studio / collage / type-only / UI surface / named mix].
• Substrate + texture: [surface, grain, print artefacts, wear]. Never a flat digital fill unless the system is Swiss or UI.
• Palette: [BG hex] background, [INK hex] for type, single accent [ACCENT hex] used sparingly, optional [SECONDARY hex] for diagrams only. Never pure black, never pure white.
• Type: [headline family + weight + case], [body family + weight], [pairing logic]. Hierarchy: [how scale and weight separate headline, body, labels].
• Signature device: [the one recurring element, described concretely: size, colour, where it sits].
• Marks: [one mark vocabulary, described concretely]. 
• Cards / diagrams: [one treatment, described vividly], used on every diagram slide.
• Image treatment: [how photos / objects enter the slide].
• No: gradient swirls, neon glow, lens flares, sparkle or lightning icons, holographic "AI" imagery, stock-photo look, invented logos, fake serial kickers ("ISSUE 01", "DISPATCH 03"), handle watermarks on every slide.
```

### Overlay table for the 9 named styles

| Style | [SYS_HEADLINE] | [SYS_BODY] | [SYS_MARK] |
|---|---|---|---|
| The Press | ultra-condensed bold grotesque sans with ONE phrase switched to Didone italic serif in [ACCENT], tilted ~3°, overlapping the sans line | clean grotesque, medium weight | one hand-drawn felt-tip mark in [ACCENT] |
| The Modernist | one sans face, heavy display weight, left-aligned on a strict grid, no italic switch, no tilt | same face, book weight, left-aligned columns | one hairline rule or one small flat geometric shape, nothing hand-drawn |
| The Dossier | serif words cut from yellowed newsprint, torn edges, each pasted at 1 to 4° | monospace typewriter | rubber-stamp date, paperclip or masking tape |
| The Billboard | one big bold display word or short phrase at the top, small tracked subtitle beneath | short caption type + one or two handwritten margin notes | washi tape on photo corners, rubber-stamp badge, colour-swatch strip |
| The Index | a gigantic numeral in [ACCENT] as the central element, single sans headline beside it | same sans, 1 to 2 short lines | hairline rule under each entry |
| The Photograph | bold sans line with ONE phrase in italic serif, over the darkened photo | clean sans over the darkened area | hand-drawn [ACCENT] scribble underline + small script signature in the opposite corner |
| The Moleskine | hand-lettered felt-tip letterforms, irregular, visible pen pressure, organic angle | monospace typewriter | several drawn marks allowed: arrows, doodled stars, scribble underlines, boxes |
| The Whiteboard | huge marker handwriting, no rendered type anywhere | smaller marker handwriting | circled words, drawn arrows, boxed notes |
| The Screenshot | the post body inside one light UI card, platform system sans, no display type outside the card | same system sans, timestamp + divider | the engagement row glyphs, nothing hand-drawn |

Using the wrong row (or writing Press clauses from memory) is the #1 way decks regress to one look. A Modernist prompt that contains "italic serif" or "hand-drawn" skipped the overlay.

### Log the locked design (free, one call)

Right after the user approves the design, BEFORE slide 1:

```
log_carousel_deck(
  composition_system="Bespoke: Cupping Sheet",   # or the named style, e.g. "The Modernist"
  palette="#F1EAD9 / #2A211B / #C8322B",         # preset name or the hexes
  mark_style="red-pen circle",
  card_treatment="score-sheet rows",
  recipe="Listicle",
  topic="7 mistakes new coffee shop owners make"
)
```

This feeds the anti-repeat memory. Fetching this skill through `get_skill(name="viral-carousel")` appends the user's recent decks; skew away from them (another source artefact, another medium, another named style). If you fetched via `resources/read` or HTTPS and see no history section, call `get_skill` once.

## Step 5: Write each slide's prompt

**The image model renders the typography INTO the image.** You are producing the finished slide, not a background for someone to add text to later. A carousel prompt that says "no text, no words, no letters" has misread this skill.

Every slide prompt contains, in this order:

1. The full DESIGN block, verbatim.
2. The archetype name and its layout, with every token filled for THIS slide (system overlay clauses included).
3. Every visible word written verbatim in quotes inside the layout description.
4. A "reads exactly" tail restating each text element: `The headline reads exactly: "..."` / `The body reads exactly: "... / ..."`. Diagrams: `The cards read EXACTLY in order, left to right: "...", "...".`
5. Concrete negative direction for this system's failure modes.
6. An image manifest when `source_media_urls` is set: one line per input, e.g. `IMAGE 1 is the slide-1 style anchor to match. IMAGE 2 is the official <brand> logo, use it exactly.`
7. Bulk microcopy WRITTEN, never described: artefact fine print (form labels, receipt lines, ingredient lists) gets 3 to 6 exact short lines plus "the rest is small illegible texture", never "dozens of tiny entries".
8. A blank-surface clause for any device, bottle or garment in frame ("the bottle is a plain unmarked frosted surface"), so no real maker's logo appears.

A correct prompt is usually 1,500 to 2,500 characters. Under 800 means the design block was cut.

### Worked example 1: bespoke system (Art director)

```
DESIGN SYSTEM (apply unchanged to every slide of this carousel):
• System: Cupping Sheet. Each slide is a coffee cupping score form printed on cream card stock, photographed flat from directly above, with notes in red ballpoint pen.
• Medium: flat-lay photograph of a printed form, real paper, soft window light from the left.
• Substrate + texture: cream uncoated card #F1EAD9 with visible fibre, faint coffee-ring stain, slightly curled corner, a pencil lying at the edge.
• Palette: card #F1EAD9, espresso ink #2A211B for printed type, red pen accent #C8322B, kraft secondary #B08A5B.
• Type: printed form headings in Druk Wide Bold, uppercase, espresso ink; form labels and scores in IBM Plex Mono; notes handwritten in red ballpoint.
• Signature device: a hand-drawn red ballpoint circle around the key word or score, slightly overshooting its start point.
• Marks: red pen only: circles, a tick, one underline. Nothing else hand-drawn.
• Cards / diagrams: rows of the score table, each row a criterion with a filled score box.
• Image treatment: no photos inside the form; the form itself is the photograph.
• No: gradient swirls, neon glow, lens flares, sparkle icons, stock coffee-cup imagery, latte art, invented logos, serial kickers, handle watermarks.

ARCHETYPE: Hook, masthead variant.

Vertical 3:4. Top-down flat lay of the cupping form on a darker walnut table edge visible at the right. Across the top of the form, huge Druk Wide Bold uppercase in espresso ink over two lines: "7 MISTAKES" / "THAT EMPTY NEW CAFÉS". Beneath, a printed form row in IBM Plex Mono: "SHOP: YOURS   SCORE: __/100". In red ballpoint, the handwritten note "we made all 7" beside the score, and a red circle drawn around "EMPTY". Below, faint printed score-table rows as small illegible texture. A coffee-ring stain overlaps the bottom-right corner.

The headline reads exactly: "7 MISTAKES THAT EMPTY NEW CAFÉS"
The form row reads exactly: "SHOP: YOURS   SCORE: __/100"
The handwritten note reads exactly: "we made all 7"
```

### Worked example 2: named style (The Modernist). Note what is ABSENT: no italic switch, no tilt, no felt-tip mark, near-flat surface.

```
DESIGN SYSTEM (apply unchanged to every slide of this carousel):
• System: The Modernist. Swiss grid, one typeface, hairline rules, generous negative space.
• Medium: type-only.
• Substrate + texture: off-white #F2EFE9, near-flat, faintest paper tone only.
• Palette: off-white #F2EFE9, ink #1A1A1A, single cobalt accent #1F47C7 (one rule or one shape per slide).
• Type: Neue Haas Grotesk in exactly two weights, heavy display for headlines, book for body. No italic switch, no tilt, everything left-aligned and baseline-aligned.
• Signature device: one small filled cobalt square anchoring the bottom-left corner of every slide.
• Marks: one cobalt hairline rule across the slide width. Nothing hand-drawn.
• Cards / diagrams: hairline-ruled table rows.
• Image treatment: photos, when used, sit in a grid cell beside the headline, never behind it.
• No: gradient swirls, felt-tip doodles, italic serif, tilted text, film grain, sparkle icons, generic tech imagery.

ARCHETYPE: Hook, left-aligned.

Vertical 3:4. Off-white background, generous negative space. A cobalt hairline rule runs full width at the top margin. Below it, left-aligned, Neue Haas Grotesk heavy display, ink, very large, three lines: "MOST SAAS / PRICING PAGES / ARE LYING." After clear vertical space, two body lines in book weight: "Not about the price." / "About who the product is for." A small filled cobalt square anchors the bottom-left corner. Nothing else.

The headline reads exactly: "MOST SAAS PRICING PAGES ARE LYING."
The body reads exactly: "Not about the price. / About who the product is for."
```

### Auto-fail prompt shape

```
BAD: "Instagram carousel slide 2. Venom Astrology brand. Dark warm near-black background. Engraved botanical illustration style. Maya astrology theme, Tzolkin glyph stone carving at center, jungle leaves, jade accents. Mood: ancient, mystical. No text, no words, no letters. Soft moody lighting."
```

No design block, no archetype, no layout, no type spec, no quoted text, "no text" (catastrophic), vibe adjectives instead of concrete references, under 600 characters. If you catch yourself writing this, stop and rebuild from the list above.

## Step 6: Generate

### Model, resolution, aspect

- **Model: `gpt-image-2-6-sunburst` for every archetype, including photo slides.** It leads on rendered typography and on edits from a reference, which is exactly what slides 2 to N are (image-to-image from the slide-1 anchor), so it holds the look across the deck. `gpt-image-2-6-flare` is its same-family sibling: same inputs, same price, faster render. Reach for Flare when turnaround on a large parallel fan-out matters more than the last bit of type and detail, or as the failover when Sunburst is down. Next fallbacks in order: `gpt-image-2-5-sunburst` (also the pick when a slide prompt runs over 8000 characters), then `nano-banana-2-1`. Slides 2 to N are edits, and a 2.6 edit adds a charge per reference image, so read the rate from `get_models` / `estimate` (pass `reference_image_count`). Do NOT fall back to `nano-banana-2` by reflex.
- `nano-banana-2-1` for a `4:5` slide only when it falls back off GPT Image 2.6 (which renders `4:5`); GPT Image 2.5 has no `4:5`, so skip it there: runner-up `nano-banana-2`, budget `nano-banana-2-lite`. Also when the user explicitly asks for Nano Banana on a documentary-photo slide and accepts weaker type on that slide. `nano-banana-pro` only on an explicit "best" / print-grade ask.
- **Resolution: `1K`** by default; it is sharp at feed sizes and costs less than 2K (read live per-resolution prices from `get_models`). Use 2K when the user asks for premium / print, or to retry a diagram whose labels misrendered at 1K. 4K only for explicit print / OOH.
- **Aspect: `3:4`** for the whole deck. GPT Image 2.6 supports `auto`, `1:1`, `3:2`, `2:3`, `4:3`, `3:4`, `5:4`, `4:5`, `16:9`, `9:16`, `2:1`, `1:2`, `21:9`, `9:21`, `3:1`, `1:3` at every resolution. Prompt cap 8000 chars on 2.6 (GPT Image 2.5 takes 20,000, and it alone has `27:16`, `16:27`, `9:8`, `8:9`, but no `4:5` or `5:4`). `source_media_urls` takes 1 to 16 images and switches to image-to-image.
- A real product, face or brand asset the user owns goes in through `kubeez://skills/media-upload` and `source_media_urls`; never imagine a product the user already has.

### Cohesion workflow

**One `generate_media` call per slide.** N swipeable slides = N calls = N images. Packing "SLIDE 1 … SLIDE 5" into one prompt produces one picture with tiny tiled panels. (A user-requested collage or grid poster is the exception: that is one image by design.)

1. Get approval on the design (Step 2) and all slide copy.
2. Generate **slide 1** alone. Show it. If the look is wrong, fix the DESIGN block (not just that prompt) and regenerate. Iterating here is cheap; fixing it after N slides is not.
3. Once slide 1 is approved, keep its `media.kubeez.com/...` URL and **fire slides 2 to N in parallel**, each on `gpt-image-2-6-sunburst` (or `gpt-image-2-6-flare` when speed matters, see above) with that URL in `source_media_urls` and this appended to the prompt: *"Match the palette, typography, texture, signature device, mark style and card treatment of IMAGE 1 exactly. Compose this slide as the [ARCHETYPE] layout of the same system."* Poll concurrently.
4. Do NOT save slides with `manage_library`; the library is for reusable brand assets, and the URL is all the next call needs.

### Known fragility (tell the user honestly)

- Text fidelity caps around **25 words per slide**. Cut copy before regenerating.
- Repeated-label diagrams are the hardest case. Labels twice in the prompt, 6 elements or fewer, retry the single slide at 2K if 1K misrenders.
- Named typefaces render far better than categories ("Playfair Display Italic" beats "italic serif").
- Marks and signature devices need concrete description (size, colour, position, gesture). "A doodle" renders nothing.
- Artefact systems fail when the fine print competes with the headline. Keep artefact text to a few exact short lines and let the rest be texture.
- UI surfaces (Screenshot, UI-native) work when stylised with short text (about 15 visible words). For a faithful replica of a real app, ask the user to upload the real screenshot and composite it via image-to-image instead of fabricating it.
- Cohesion reaches about 85% with the slide-1 anchor; regenerate one or two outliers individually.

## Step 7: Review before shipping

Run the per-slide critique and the deck-level cohesion check from `kubeez://skills/critique`. Per slide: text renders exactly, headline reads at thumbnail size, the signature device and palette are present, no forbidden elements. Deck level: swipe through in order; every slide looks like the same object. Regenerate misaligned slides individually, never the whole deck.

## Anti-patterns (auto-fail)

- Background + centered text, brand-colour background + centered text, plain gradient behind a subject.
- A "bespoke" system that is a library style with new colours, or that fails the swap test.
- Austere gallery-art hooks (one word on blank paper, zero-text photo) for non-luxury topics.
- Mixing systems, palettes, card treatments or mark vocabularies mid-deck.
- Writing Press-style clauses (condensed sans + italic serif + felt-tip) into a deck that locked another system.
- Invented numbers, quotes or dates; fake serial kickers; a handle watermark on every slide (handle only on request, CTA slide only).
- Sparkle / lightning / rocket icon glyphs, neon glow, lens flares, holographic or circuit-board "AI" imagery. (A hand-drawn star doodle inside a hand-drawn mark vocabulary is fine; an icon glyph is not.)
- Fabricated real-brand UI passed off as authentic, or dense real-app chrome with hallucinated fine print.
- Defaulting to 10 slides, to the same named style as the last deck, to 2K / 4K "to be safe", or to `nano-banana-2` by reflex.
- Short vibe-only prompts, or "no text, no words, no letters" in a slide prompt.

## Pre-generation checklist

- [ ] Freshness check done where the topic needs it; sources cited; factual brief approved.
- [ ] Mode picked (Reference / Named style / Art director). Gallery offered at most once.
- [ ] Art director: source artefact chosen, two concepts drafted and one kept, every spec axis filled, all four tests passed.
- [ ] Named style: library fetched; palette, cards, type pairing and mark vocabulary locked; overlay row copied.
- [ ] Slide count from topic depth; archetype per slide; no two adjacent layouts the same; hook headline 8 words or fewer; every slide 25 words or fewer.
- [ ] Design + copy approved by the user; `log_carousel_deck` called before slide 1.
- [ ] DESIGN block ready to paste verbatim; each prompt clears the Step 5 list (quoted text, "reads exactly" tail, image manifest, blank-surface clause).
- [ ] `gpt-image-2-6-sunburst` (Flare only for speed), `1K`, `3:4`, unless an exception above applies.
- [ ] Slide 1 alone first; slides 2 to N in parallel with the slide-1 URL as `source_media_urls`; no `manage_library` saves.
- [ ] Critique + cohesion check before shipping.
