Carousel Workflows

The carousel workflow generates an ordered deck of slides for a single-topic carousel post. The LLM drafts a per-slide outline (slide 1 = hook, last slide = CTA, middles = distinct points), then one image is generated per slide using a prompt derived from that slide’s outline entry. The output preserves slide order so consumer UIs can render the deck as a swipeable group rather than a grid of unordered thumbnails.

Tip

Use carousel/generate when you want a single coherent deck (slide 1 is the hook, slide N is the CTA, all slides share visual style). Use image/generate when you want N independent thumbnail variants of the same hook for selection.

carousel/generate

Workflow: carousel/generate  ·  Aliases: carousel, image/carousel

Produces a single carousel post — outline + N rendered slides in narrative order, plus a post-body caption.

Submit either the canonical name or the image/carousel alias (Socialite uses the latter from its bundle dispatcher so it can swap from image/generate without changing the input shape).

Quick Start

CLI

fab run carousel/generate \
  --input topic="AI productivity tips" \
  --input slide_count=5 \
  --input audience="busy founders" \
  --input tone="punchy"

REST

curl -X POST "$FABRIC_URL/v1/workflows/run?name=carousel/generate" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "topic": "AI productivity tips",
      "slide_count": 5,
      "audience": "busy founders",
      "tone": "punchy",
      "channels": ["instagram"],
      "aspect_ratio": "1:1"
    }
  }'

TypeScript SDK

import { FabricClient } from "@fabric-platform/sdk";

const fabric = new FabricClient({ apiKey: process.env.FABRIC_API_KEY });

const run = await fabric.workflows.runs.submit("carousel/generate", {
  input: {
    topic: "AI productivity tips",
    slide_count: 5,
    audience: "busy founders",
    channels: ["instagram"],
  },
});

Pipeline

draft_outline (LLM)  →  derive_aesthetic (LLM)  →  render_slides (image-per-slide, parallel)  →  write_caption (LLM)

1. draft_outline asks the LLM for a [{title, body}, ...] plan with slide_count entries. Slide 1 is the hook; the last slide is the CTA. Titles are short (≤ 6 words) and ALL CAPS-ready.
2. derive_aesthetic drafts a single concrete aesthetic brief (palette, lighting, composition, medium) for the whole deck so every slide reads as part of the same set. Skipped when the caller passed an explicit visual_style.
3. render_slides generates one image per outline entry in parallel. Each prompt combines the slide title (on-image hero text), the deck-wide aesthetic, and tone-driven visual directives (e.g. “punchy” → saturated complementary colors, hard rim lighting, dynamic composition).
4. write_caption drafts the post-body caption (2–4 sentences, soft CTA) referencing the realized slide titles, so the caption matches what the deck actually says.

Input

Parameter	Type	Default	Description
topic	string	required	Subject of the carousel — drives outline + image prompts.
slide_count	int	5	Number of slides (3–10).
num_thumbnails	int | null	null	Legacy alias for slide_count (accepted so callers swapping from image/generate don’t have to rename the field).
audience	string	"general"	Target reader persona.
tone	string	""	Voice/tone hint (e.g. "punchy", "warm", "luxurious"). Drives both narrative voice and the image model’s visual direction (palette, lighting, energy).
industry	string	""	Industry/vertical hint passed through to outline + caption prompts.
keywords	list[str]	[]	Topical keywords woven into the deck’s narrative.
channels	list[str]	[]	Distribution channels; shapes prompt guidance.
visual_style	string	""	Optional override. When empty, the workflow derives a deck-wide aesthetic brief from topic + tone + audience (recommended).
aspect_ratio	string	"1:1"	Aspect ratio for slide images. Imagen accepts 1:1, 9:16, 16:9, 4:3, 3:4. Avoid 4:5 — Imagen rejects it and the run silently falls through to a lower-quality provider.

Output

{
  "kind": "carousel",
  "topic": "AI productivity tips",
  "caption": "AI won't take your job — but the founder using it will. Swipe through 5 ways to claw back two hours a day, starting with the meeting you can already kill.",
  "slides": [
    {
      "order": 0,
      "title": "AI WON'T REPLACE YOU",
      "body": "But someone using it will.",
      "image_path": "/tmp/carousel_slide_00.png",
      "image_asset_id": "018f..."
    },
    {
      "order": 1,
      "title": "TIP 1 — DELETE BUSYWORK",
      "body": "Automate the meetings nobody asked for.",
      "image_path": "/tmp/carousel_slide_01.png",
      "image_asset_id": "018f..."
    }
  ]
}

Field	Type	Description
kind	string	Always "carousel" — consumers route on this.
topic	string	Echoed input topic.
caption	string	Post-body caption summarizing the deck — what the consumer renders as the Instagram post copy under the swipeable images.
slides	list[CarouselSlide]	Ordered slide list. slides[i].order matches the index.
slides[].order	int	Zero-based slide position. Stable across the run.
slides[].title	string	On-image hero text.
slides[].body	string	Optional supporting copy (may be empty for visual-only slides).
slides[].image_path	string	Local or signed-URL path to the rendered slide.
slides[].image_asset_id	string | null	Fabric asset ID when uploaded server-side.

Failure semantics

If a single slide’s image generation fails, the slide is dropped from the output but the run still succeeds. If every slide fails, the run fails with a clear error. Consumers should iterate slides (which may have gaps in order) rather than assuming a contiguous index.

Use with the bundle dispatcher

Pair carousel/generate with bundle/ad-hoc to produce a multi-format deck — carousel + video + text post — for the same brief in one submission:

{
  "bundle": [
    { "workflow": "carousel/generate", "input": { "topic": "...", "slide_count": 5 } },
    { "workflow": "video/ai-shorts", "input": { "topic": "..." } },
    { "workflow": "global/content-generate", "input": { "topic": "...", "content_types": ["text"] } }
  ]
}