Carousel

Trigger

FIRES on a generic multi-slide carousel brief: "make an IG carousel about X", "5-slide LinkedIn post", "swipe-through deck", "story-style series of N slides", "multi-style carousel with 6 aesthetics", "explainer carousel for our launch". Any subject — the skill is subject-agnostic.

DO NOT FIRE when:

• The brief is for ONE still (poster, drop graphic) → use /poster instead.
• The brief is for N static ads across registers (FB / Meta ad pack) → use /fb-creatives instead.
• The brief is for a video (the slides animate, the mascot moves) → match a /ralphy-ugc-* video skill + editor playbook.
• The user points at one specific carousel to reproduce → that is the remix path. See docs/skills-vs-templates.md.

What this skill is

A generalized cover-first / dual-ref / mascot-fit overlay, not a finished carousel. It does not name a mascot, a topic, or an aesthetic — it tells the art-director HOW multi-style carousels are built so a strong one comes out for whatever the user is pushing. It runs through ralphy generate image; no HyperFrames compose stage is needed when the text bakes inside the image.

Hard invariants

• All generation routes through ralphy generate image (no raw API). Read MODELS.md before naming any model id.
• gpt-5.4-image-2 for baked text on carousel slides. Validated 30/30 slides on ralphy-carousel-001 for 1-2-line headlines + small label + sub-line legible and spelled right. gemini smudges small embedded type — don't finalize there.
• Carousel batches serialize per style, but parallel across styles is fine. Don't run two bg gen-loops on the same OpenRouter key writing to the same prompt files at once.
• Never mutate prompt-input files while a background gen-loop is reading them. The gen loop cats each prompt file per iteration — a mid-flight rm or mv corrupts the batch. Edit / delete prompt files AFTER background jobs finish (or before they start). See ralphy-carousel-001 Finding B.
• Append-only on regen. Re-rolling a slide writes .v2.png; the pre-fix punk-01.v1.png and acid-01.v1.png stay on disk for A/B.
• The reference-required gate fires when the brief names a real person / branded product / IP in a slide subject. A no-name mascot + a fictional brand proceeds without a ref.
• The quality gate refuses, not warns — two failed scoreImage in a row → stop and report options.

The niche, in one paragraph

A multi-style carousel lives or dies on per-style cohesion + cross-style identity. Each 5-slide set must read as one design language (same paper / light / grade / type system), AND the mascot or brand must read as the same entity across every set. Two refs glued together do this work: a permanent mascot-ref.png anchors identity across the whole deck; the approved cover for each style anchors that style's internal cohesion across slides 02-05. A clean cute mascot doesn't survive grunge — for distressed registers, build a dedicated character variant rather than forcing the clean asset.

The fixed workflow (do not restructure — only fill the slots)

The cover-first checkpoint is the hard gate. Catching a mascot-fit failure here costs 1 cover ($0.20); catching it after the full set burns ~$1.60 of blind gens.

The dual-ref cohesion rule

Every fill-slide gen passes BOTH refs:

ralphy generate image --project <id> --slot <style>-NN \
  --model openai/gpt-5.4-image-2 --size 1080x1350 \
  --ref artifacts/refs/mascot-ref.png \
  --ref artifacts/images/<style>-01.png \
  --prompt "$(cat prompts/<style>-NN.json)"

• mascot-ref.png holds the mascot's identity across the deck (no drift in geometry / color / character).
• <style>-01.png (the approved cover) holds the style's internal cohesion (same paper / light / grade / type) across slides 02-05.

Without both, either the mascot drifts across styles OR the style drifts across slides — usually both.

The mascot-fit rule

A clean / cute / friendly brand mascot does not survive gritty registers (xerox photocopy, halftone, grunge, acid). Forcing the clean asset into them produces a "sticker pasted on a poster" mismatch — the user will reject it.

Two valid moves:

1. Reinterpret in the medium (passable). Prompt-token verbatim: "the REDRAWN ENTIRELY IN THE POSTER'S OWN MEDIUM — a 1-bit photocopied duotone screen-print … rough torn edges … NOT a clean 3D object, NOT a glossy sticker cut-out — it must look printed and distressed, fully part of the page".
2. Dedicated distressed character variant (better). Build artifacts/refs/mascot-gritty-ref.png once; pass it instead of the clean ref on punk / acid / xerox slides.

Decide at the cover-first checkpoint, not after the full set.

JSON prompt schema (one STYLE + QUALITY block per aesthetic)

{
  "scene": "<the per-slide subject — varies per slide>",
  "style": "<reused per style — paper, light, grade, type system, accent>",
  "technical": "<reused per style — printing medium, halftone settings, grain>",
  "composition": {
    "ui_elements": [
      "<headline text baked on slide, exact spelling>",
      "<small label baked on slide>",
      "<one sub-line baked on slide>"
    ]
  },
  "quality": "<reused per style — fidelity tokens + negatives>"
}

Locking style + technical + quality per aesthetic and only swapping scene + ui_elements per slide is what makes a 5-slide set read as one design.

Single-accent lock

In every prompt body, name the ONE accent color in hex AND add the source hue to AVOID. Example for a brand-orange recolor: bright Ralphy-ORANGE #FFA630 — NOT green in description + "green color" in the AVOID list. Without this, gpt-image drifts into rainbow chrome and the style identity collapses.

Default model stack (verify against MODELS.md)

• Default — openai/gpt-5.4-image-2. Best at baked text (30/30 slides spelled right). Serialize per style (1 concurrent per key for the loop on that style); parallel across styles is fine. Aspect via --size 1080x1350 (4:5 IG / LinkedIn carousel default), --size 1080x1080 for 1:1, --size 1080x1920 for Stories.
• Fallback — google/gemini-3-pro-image-preview. Faster, but smudges small embedded typography. Use ONLY for fast palette / cover exploration; finalize on gpt-image.

Failure modes

• Forcing a clean mascot into gritty styles. Cost on ralphy-carousel-001: $0.40 cover re-rolls + weaker result. Prevention: cover-first checkpoint → reinterpret-in-medium OR dedicated distressed character ref.
• rm prompts/<slot>.txt during a running bg loop. The loop cats lazily — delete corrupts the batch. Prevention: edit / delete prompt files only after bg jobs finish.
• Transient TLS / socket exit-1 mid-batch. ~10 min sunk on ralphy-carousel-001 (3 slots). No auto-retry in CLI today. Prevention: re-run only the affected slot; the rest of the batch is intact.
• Abandoning a style direction after slide-02. Decide direction before generating past one proof. Prevention: cover-first checkpoint is the decision gate.
• Trying to fan out 2 gpt-image loops in parallel on one key → 403. Serialize per style; parallel across styles uses different prompts but only as many concurrent calls as the key tolerates (probe-test before fanning out beyond 1 / style).
• "Use our colors" missing the AVOID hue. The model defaults toward the source aesthetic's hue; explicit AVOID negates it.

Aspect ratio

• IG carousel default → --size 1080x1350 (4:5).
• LinkedIn document-style carousel → same or --size 1080x1080 (1:1).
• TikTok / Stories carousel → --size 1080x1920 (9:16).

CLI cookbook

# Step 2 — rasterize mascot SVG into a ref PNG (use playwright or a designed render).
# Output: artifacts/refs/mascot-ref.png (and artifacts/refs/mascot-gritty-ref.png for distressed styles).

# Step 4 — cover-first checkpoint — fan out one cover per style in parallel.
for style in zine club swiss riso punk acid; do
  ralphy generate image --project <id> --slot "${style}-01" \
    --model openai/gpt-5.4-image-2 --size 1080x1350 \
    --ref artifacts/refs/mascot-ref.png \
    --prompt "$(cat prompts/${style}-01.json)" &
done
wait
# Show all covers in chat, get user approval per style. Re-roll only mismatches.

# Step 5 — fill 02-05 per approved style (serialized inside a style, parallel across styles).
for style in zine club swiss riso; do
  (
    for slide in 02 03 04 05; do
      ralphy generate image --project <id> --slot "${style}-${slide}" \
        --model openai/gpt-5.4-image-2 --size 1080x1350 \
        --ref artifacts/refs/mascot-ref.png \
        --ref "artifacts/images/${style}-01.png" \
        --prompt "$(cat prompts/${style}-${slide}.json)"
    done
  ) &
done
wait

# Step 6 — contact sheet per style (one row of 5 slides)
for style in zine club swiss riso punk acid; do
  ffmpeg -i "artifacts/images/${style}-01.png" -i "artifacts/images/${style}-02.png" \
         -i "artifacts/images/${style}-03.png" -i "artifacts/images/${style}-04.png" \
         -i "artifacts/images/${style}-05.png" \
         -filter_complex "hstack=inputs=5" "contact/${style}-row.png"
done

# Re-roll a single transient-network failure
ralphy generate image --project <id> --slot riso-01 \
  --model openai/gpt-5.4-image-2 --size 1080x1350 \
  --ref artifacts/refs/mascot-ref.png \
  --prompt "$(cat prompts/riso-01.json)"

At ~$0.20 / slide on gpt-image, a 6-style × 5-slide carousel = ~$6.00 minimum (the ralphy-carousel-001 reference ran $6.80 / 1.15×).

See also

• docs/skills-vs-templates.md — why this is a skill and not a template.
• .agents/skills/art-director/SKILL.md — ref-anchor flow and model picks.
• .agents/skills/intake/SKILL.md — the multi-slide branch.
• MEMORY.md — append-only-on-generations, anti-ai-slop image prompts.
• Reference postmortem: .ralphy/workspaces/<ws>/projects/ralphy-carousel-001/postmortem/ — the cover-first / dual-ref / mascot-fit rules this skill codifies.