Generate Carousel

generate-carousel

The full pipeline for producing a multi-slide Instagram carousel — research, design, generate, host, caption, publish.

When to invoke

• The user wants a polished Instagram carousel for a brand or creator account
• They want every slide generated with one image model (gpt-image-2) including the type, not photoshopped or designed in Figma
• The post needs to look native to a category — listicle, format comparison, use-case showcase, "this is X" reveal, or any standard 3–10 slide structure

Required environment variables

• WAVESPEED_API_KEY — required, image generation
• SCRAPE_CREATORS_API — required for niche research (pull reference posts to identify the format)
• A CDN with a CLI or REST endpoint that returns a public URL — required for hosting the PNGs the publisher pulls from. Common choice: VidJutsu (vidjutsu upload <file> returns a cdn.vidjutsu.ai/... URL)
• ZERNIO_API_KEY — required to publish or schedule via the social publisher
• GEMINI_API_KEY — optional, for 4K via Nano Banana 2 if a slide demands it

The pipeline (7 steps)

Read each recipe in order. Every recipe is a self-contained markdown file under recipes/.

1. Research the format → recipes/01-research-formats.md
2. Design the carousel → recipes/02-design-the-carousel.md
3. Write slide specs → recipes/03-write-slide-specs.md
4. Generate slides → recipes/04-generate-slides.md
5. Host on a CDN → recipes/05-host-slides.md
6. Draft the caption → recipes/06-write-caption.md
7. Publish → recipes/07-publish.md

Working directory layout

Set up the project's working directory exactly as described in references/folder-structure.md before any step.

Schema and endpoint references

• references/folder-structure.md — canonical project layout
• references/format-grammar.md — the recurring visual formats carousels use, and how to spot the right one
• references/slide-spec.md — JSON schema for a single slide
• references/typography-rules.md — how to bake type into a gpt-image-2 prompt so the model renders headline, logo lockup, and CTA pill in-image
• references/scrapecreators-instagram.md — endpoints for pulling profile + posts + individual carousel slides
• references/wavespeed-gpt-image-2.md — image generation endpoint and gotchas
• references/vidjutsu-cdn.md — uploading PNGs for a public URL
• references/zernio-instagram.md — creating a carousel post
• references/caption-format.md — caption structure that pairs with a carousel

Style guardrails

• One image model for everything. gpt-image-2 generates the photograph AND the typography in a single pass. Do not design the type separately and overlay it.
• Typography rules in references/typography-rules.md are load-bearing — follow them to avoid garbled text.
• Aspect ratio: 4:5 portrait (Instagram carousel native).
• Resolution: 2K from gpt-image-2 unless a slide requires 4K (Nano Banana 2).
• Visual continuity across the carousel: same logo lockup placement, same type system, same single-italic-word motif if the brand uses one. Variation comes from the photography, not the type.
• Caption pairs with the cover slide hook and ends with a single clear CTA that matches what the closer slide promises.

Cost expectations

• One slide at 2K via Wavespeed gpt-image-2 ≈ ~$0.02
• A 7-slide carousel ≈ ~$0.14 in image-gen credits, plus ~free CDN hosting and ~free publisher
• Total time end to end (with research): 30–90 minutes