Newsjack Carousel

When to invoke

• The user shares a URL to a launch, news drop, model release, viral post, product reveal, or any topical event and wants a carousel ABOUT it for their account
• They want to react to or announce a third party's news (e.g., "Higgsfield just dropped MCP — make a carousel announcing it")
• The source post is video, reel, carousel, or single image — the skill adapts to all four
• They optionally provide a second URL as the visual format reference (e.g., "match the format of this carousel")

This is the topical / newsjack sibling of clone-carousel. Difference:

• clone-carousel — input MUST be a carousel; output reuses its FORMAT for the user's own brand niche (no relation to the source's subject matter)
• newsjack-carousel — input can be ANY post; output is ABOUT the source's subject matter (the source is the topic seed, not the format seed); a second optional URL provides the format

If the user wants to clone a format and keep it niche-agnostic, use clone-carousel. If the user wants to react to / announce a specific topic from a URL, use this skill.

Required environment variables

• SCRAPE_CREATORS_API — required. Pulls the topic post (and optional format reference).
• GEMINI_API_KEY — required. Analyzes frames + caption to determine the announcement, key facts, and visual style.
• WAVESPEED_API_KEY — required. Slide generation via gpt-image-2.
• VIDJUTSU_API_KEY — required. CDN hosting for stable public slide URLs.
• ZERNIO_API_KEY — required. Publishes to the user's connected account.

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

The pipeline (6 steps)

1. Fetch the topic post → recipes/01-fetch-topic.md — Scrape Creators handles video / reel / carousel / image. Downloads media + caption + author handle.
2. Analyze the topic → recipes/02-analyze-topic.md — Gemini reads frames (or carousel slides) + caption and returns structured JSON: announcement summary, key features, visual style, and a one-line carousel angle.
3. Optionally fetch a format reference → recipes/03-pick-format.md — if the user provided a second URL, pull and analyze it the same way clone-carousel does. Otherwise pick a default tech-announcement format (black bg, brand-color heading, white body, photo card lower half).
4. Write slide specs → recipes/04-write-slide-specs.md — 6–8 slides in the user's brand voice: cover (announcement), 4–6 capability/proof slides drawn from the analysis, CTA closer. Brand color is the user's, NOT the source's — match Higgsfield green for Higgsfield news, Anthropic orange for Claude news, etc.
5. Generate, host, draft caption → recipes/05-generate-and-host.md — Wavespeed gpt-image-2 in parallel → VidJutsu CDN upload → caption drafted in the user's voice mirroring the format reference's caption pattern.
6. Confirm and publish → recipes/06-publish.md — Surface payload (caption + slide order + target account + schedule) for approval. Fire POST /v1/posts with publishNow: true. Stop polling on the permalink — surface the post ID and let Zernio finalize asynchronously.

Working directory layout

CWD/newsjack-carousel-<YYYYMMDD-HHMM>/
├── topic/
│   ├── post_<shortcode>.json       # Scrape Creators full body
│   ├── reference.mp4               # if video/reel
│   ├── frames/frame_*.jpg          # extracted via ffmpeg fps=1/3
│   └── analysis.json               # Gemini structured output
├── format/                         # only if user gave a 2nd URL
│   ├── post_<shortcode>.json
│   └── images/slide_*.jpg
├── specs/
│   └── slide_<N>_<role>.json       # one per slide
├── frames/
│   └── slide_<N>_<role>.png        # final 4:5 2K renders
├── caption.md
├── publish_payload.json
└── post.md                         # zernio post id + permalink + slide URLs

Schema and endpoint references

• references/slide-spec.md — JSON schema for one slide
• references/typography-rules.md — baking type into gpt-image-2 in a single pass
• references/wavespeed-gpt-image-2.md — endpoint, params, parallelism, content-flag retry
• references/vidjutsu-cdn.md — uploading PNGs for stable public URLs
• references/zernio-publish.md — create-post payload, the publishNow gotcha, stuck-publish recovery
• references/scrapecreators-instagram.md — post-fetch endpoint shapes for video, reel, carousel, image

Style guardrails

• Brand color belongs to the user, not the source. If the topic post is from Higgsfield (neon lime), the user posts in their own brand color. Pull the user's brand color from MEMORY.md or ask once and remember.
• Announcement framing. The cover headline names the topic + drops the news. The body slides are concrete value (features, prompts to try, before/after, proof points). The closer is a CTA.
• Source attribution in the caption opening line. Lead with @<source-handle> just shipped X — gives credit, signals topical relevance, and improves IG ranking via the @-mention.
• No reproduction of source visuals. The carousel is ABOUT the topic; never reuse the source's photographs, logos, or screen recordings beyond a small attribution mark if the user explicitly asks.
• Polling discipline. Per user preference: do not start tight poll loops on Zernio/Wavespeed/Meta without surfacing the cadence and total duration first. Default to ScheduleWakeup at 2–10 min intervals over short-sleep loops.
• Confirmation before publish. Surface the full payload (caption, slide order, target @username, schedule). Wait for explicit go.

Cost and time expectations

• Scrape Creators: ~$0.01 per post (×1 if no format ref, ×2 if user provided one)
• Gemini 2.5 Pro analysis: ~$0.02–0.05 per analysis (depends on # of frames passed)
• Wavespeed gpt-image-2 at 2K, 4:5: ~$0.02 per slide
• VidJutsu CDN: free
• Zernio publish: free

A typical 8-slide newsjack carousel: ~$0.20 in API costs, ~6–10 min wall time, including a single regen on one slide if typography misfires.

Failure modes worth knowing upfront

• Source URL is a Reel/video and user asked clone-carousel — wrong skill; this one handles it.
• Source is a paywalled or private post — Scrape Creators returns 404; tell user and stop.
• Gemini SSL cert error from system Python — use curl via subprocess instead of urllib.request. See recipes/02-analyze-topic.md.
• Zernio dedup error after a stuck publish — DELETE the stuck post first, tweak the caption (one sentence is enough), then re-POST. Don't try to PUT-then-publish — it stays as a draft.
• Brand color mismatch. If user has not set a brand color, ask once and save to memory; don't guess.