Clone Carousel

clone-carousel

The full pipeline for cloning a winning Instagram carousel — fetch reference, analyze format, write specs, generate slides with baked typography, host on a CDN, caption, and publish.

When to invoke

• The user has a specific Instagram carousel URL they want to clone the format of for their own account
• They want every slide generated by one image model (gpt-image-2) including the typography — no Photoshop, no Figma, no overlay step
• The reference is a public Instagram post URL with at least 2 slides (single-image posts are out of scope; use a different skill)
• They have a connected Instagram account in Zernio and are ready to publish or push to drafts

This is the carousel sibling of clone-ad. It clones format, not photos. If the user wants original work researched from scratch in a niche, use generate-carousel instead.

Required environment variables

• SCRAPE_CREATORS_API — required. Used to pull the reference Instagram post and download every carousel slide.
• WAVESPEED_API_KEY — required. Image generation via gpt-image-2.
• VIDJUTSU_API_KEY — required. CDN hosting via vidjutsu upload for stable public slide URLs the publisher can fetch.
• ZERNIO_API_KEY — required. Publishes the carousel to the user's connected Instagram account.

The skill checks all four before any provider call. If a key is missing it stops with a signup link.

The pipeline (6 steps)

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

1. Fetch the reference post → recipes/01-fetch-reference.md — Scrape Creators → caption + carousel children → download every slide locally
2. Analyze the format → recipes/02-analyze-format.md — read slides, write the format rules: cover does X, middle slides hold Y, closer does Z
3. Write slide specs for the brand → recipes/03-write-slide-specs.md — same format, fresh scenes, structured JSON one per slide
4. Generate slides → recipes/04-generate-slides.md — Wavespeed gpt-image-2 in parallel with safety-flag retry handling
5. Host on a CDN → recipes/05-host-slides.md — vidjutsu upload → stable cdn.vidjutsu.ai URLs
6. Caption and publish → recipes/06-caption-and-publish.md — Zernio create-post with publishNow: true and mediaType: CAROUSEL_ALBUM

Working directory layout

Outputs land in CWD/clone-carousel-<timestamp>/. See references/folder-structure.md.

Schema and endpoint references

• references/folder-structure.md — canonical project layout
• references/scrapecreators-instagram.md — the post-fetch endpoint and carousel-children parsing
• references/format-rules.md — how to read a reference carousel and write the format rules an LLM can act on
• references/slide-spec.md — JSON schema for a single slide (subject, environment, camera, lighting, typography string)
• references/typography-rules.md — how to bake type into a gpt-image-2 prompt so headline, logo lockup, and CTA pill render in-image without garbling
• references/wavespeed-gpt-image-2.md — endpoint, params, parallelism, content-moderation retry pattern
• references/vidjutsu-cdn.md — uploading PNGs for a stable public URL
• references/zernio-publish.md — create-post payload, the publishNow gotcha, draft vs. immediate publish, regenerating a single slide on a draft

Style guardrails

• One image model for everything. gpt-image-2 generates the photograph AND the typography in a single pass. No overlay step, no Photoshop, no Figma.
• Clone the format, not the photos. The reference is for structure (cover rule, middle-slide rule, closer rule). Substitute scenes for the user's brand. Never reproduce a reference image.
• Aspect ratio: 4:5 portrait at 2K. Native Instagram carousel size.
• Typography continuity across slides. Same color, same weight system, same single-italic-word motif if the reference uses one. Variation comes from the photograph, not the type.
• Caption mirrors the reference's structure, not its words. Same hook shape, same payoff length, same CTA mechanic — fresh content.
• Confirm before publishing. Surface the Zernio request body for user approval before firing the create-post call. Default to draft (far-future scheduledFor) if the user hasn't said go.

Cost and time expectations

• Scrape Creators: ~$0.01 per reference fetch
• Wavespeed gpt-image-2 at 2K, 4:5: ~$0.02 per slide
• VidJutsu CDN upload: free per file
• Zernio publish: free per post

A 7-slide clone end-to-end: ~$0.15 in API costs, ~5–10 min wall time including a single regeneration loop on one slide.

Budget for one re-render if a slide gets flagged by gpt-image-2's safety filter (rewrite the spec with safer wording and resubmit) or if typography on the cover or closer renders garbled. Both are cheap and quick.