Composed Workflows

Composed workflows chain simpler workflows into end-to-end pipelines. They combine research, hook generation, and video production into single invocations.

Research to Shorts

Workflow: video/research_to_shorts

Deep research on a topic, generate viral hooks from findings, select the best hook, and produce a complete AI short video — grounded in real data.

CLI
REST

fabric run video/research_to_shorts \
  --input query="Why developers are switching to Rust" \
  --input platform="TikTok" \
  --input quality=premium

curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/research-to-shorts" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "query": "Why developers are switching to Rust",
      "platform": "TikTok",
      "quality": "premium"
    }
  }'

Pipeline

deep_research → bridge_research_to_hooks → generate_hooks →
select_best_hook → generate_script (research-grounded) →
video_production_pipeline

Input

Parameter	Type	Default	Description
`query`	`string`	required	Research topic
`platform`	`string`	`"TikTok"`	Target platform
`duration_secs`	`int`	`45`	Video duration
`mood`	`string`	`"high-energy"`	Visual and narrative mood
`quality`	`string`	`""`	Quality preset

Output

Complete video with research-informed script and all intermediate artifacts (research synthesis, hooks, script, b-roll paths).

Hot Topics to Short

Workflow: composed/hot_topics_to_short

Scan trending topics, generate hooks, pick the strongest, and produce a single viral short video.

CLI
REST

fabric run composed/hot_topics_to_short \
  --input platform="YouTube Shorts" \
  --input categories='["tech", "AI"]'

curl -X POST "$FABRIC_URL/v1/workflows/run?name=composed/hot-topics-to-short" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "platform": "YouTube Shorts",
      "categories": ["tech", "AI"]
    }
  }'

Pipeline

hot_topics_pipeline → bridge_topics_to_hooks → generate_hooks →
select_best_hook → generate_script → video_production_pipeline

Hot Topics to Content

Workflow: composed/hot_topics_to_content

Scan trending topics, generate hooks, and batch-produce content items (scripts, outlines, or posts).

CLI
REST

fab-workflow composed/hot_topics_to_content \
  --input categories='["tech", "business"]' \
  --input num_content=5

curl -X POST "$FABRIC_URL/v1/workflows/run?name=composed/hot-topics-to-content" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "categories": ["tech", "business"],
      "num_content": 5
    }
  }'

Pipeline

hot_topics_pipeline → bridge_topics_to_hooks → generate_hooks →
generate_content_batch

Output

{
  "content_items": [
    {
      "topic": "...",
      "hook": "...",
      "hook_type": "controversial_take",
      "estimated_strength": 0.87
    }
  ]
}

Hot Topics to Content with Text Adaptation

Workflow: composed/hot_topics_to_content_with_text

Extended version that adds cross-platform text adaptation after content batch generation. Produces both content items and platform-native text for the top hook.

CLI
REST

fab-workflow global/hot-topics-to-content-with-text \
  --input categories='["tech", "AI"]' \
  --input platforms=instagram,linkedin,twitter,threads \
  --input tone_mode=narrative

curl -X POST "$FABRIC_URL/v1/workflows/run?name=composed/hot-topics-to-content-with-text" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "categories": ["tech", "AI"],
      "platforms": ["instagram", "linkedin", "twitter", "threads"],
      "tone_mode": "narrative"
    }
  }'

Pipeline

hot_topics_pipeline → bridge_topics_to_hooks → generate_hooks →
generate_content_batch → bridge_content_to_platform_adapt →
adapt_for_platforms

Output

Includes everything from hot_topics_to_content plus:

{
  "platform_texts": [
    {
      "platform": "instagram",
      "text": "Caption text adapted for Instagram...",
      "hashtags": ["ai", "tech", "productivity"],
      "char_count": 142,
      "tone_mode": "narrative"
    },
    {
      "platform": "linkedin",
      "text": "Professional post adapted for LinkedIn...",
      "hashtags": ["artificialintelligence", "techleadership"],
      "char_count": 847,
      "tone_mode": "narrative"
    }
  ]
}

Research to Video (Fan-Out)

Workflow: composed/research_to_video

Research a topic, generate hooks, then fan out N parallel AI Shorts runs — one per hook. Produces multiple videos from a single research session.

CLI
REST

fabric run composed/research_to_video \
  --input topic="AI replacing junior developers" \
  --input videos=3 \
  --input quality=premium

curl -X POST "$FABRIC_URL/v1/workflows/run?name=composed/research-to-video" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "topic": "AI replacing junior developers",
      "videos": 3,
      "quality": "premium"
    }
  }'

Pipeline

deep_research → generate_hooks → select_top_N →
fork(ai_shorts × N)

Input

Parameter	Type	Default	Description
`topic`	`string`	required	Research topic
`videos`	`int`	`3`	Number of videos to produce
`quality`	`string`	`""`	Quality preset for all videos
`--from-research`	`string`	—	Reuse prior research output file
`--from-hooks`	`string`	—	Reuse prior hooks output file

Research to Hooks

Workflow: composed/my_research_hooks

Lightweight pipeline: research a topic and extract viral hooks. No video production.

CLI
REST

fabric run composed/my_research_hooks --input query="AI agents in production"

curl -X POST "$FABRIC_URL/v1/workflows/run?name=composed/my-research-hooks" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "query": "AI agents in production"
    }
  }'

Pipeline

deep_research → bridge_to_hooks → generate_hooks

Output

{
  "hook_ideas": [
    {
      "hook_text": "Most AI agents fail in production. Here's why.",
      "hook_type": "controversial_take",
      "emotional_trigger": "fear_of_failure",
      "estimated_strength": 0.91
    }
  ]
}

Epic Remix

Workflow: composed/epic_remix

Takes a YouTube, TikTok, or Instagram video, detects its most epic moments, extracts the key ideas, then generates a completely original short-form video inspired by those ideas — styled to a user-defined persona. No source footage is reused.

CLI
REST

# With a style profile JSON file
fabric run composed/epic_remix \
  --input url="https://youtube.com/watch?v=..." \
  --input style_profile='{"tone":"sarcastic and irreverent","persona":"tech skeptic who has seen it all"}'

curl -X POST "$FABRIC_URL/v1/workflows/run?name=composed/epic-remix" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "url": "https://youtube.com/watch?v=...",
      "style_profile": {
        "tone": "sarcastic and irreverent",
        "persona": "tech skeptic who has seen it all"
      }
    }
  }'

# CLI shorthand
python -m workflows.composed.epic_remix "https://youtube.com/watch?v=..." \
  --tone "sarcastic and irreverent" \
  --persona "tech skeptic who has seen it all" \
  --hooks-style "rhetorical questions" \
  --visual-aesthetic "warm film grain"

# Ideas only (no video generation)
python -m workflows.composed.epic_remix "https://tiktok.com/@user/video/123" \
  --tone "warm and educational" --persona "excited science nerd" \
  --ideas-only

Pipeline

transcription → detect_viral_moments → extract_epic_ideas →
generate_remix_script (style-aware) → video_production_pipeline

Input

Parameter	Type	Default	Description
`url`	`string`	required	YouTube, TikTok, or Instagram video URL
`style_profile`	`object`	required	Style profile (see below)
`platform`	`string`	`"TikTok"`	Target platform
`duration_secs`	`int`	`45`	Video duration
`quality`	`string`	`"premium"`	Quality preset (`cheap`, `premium`, `ultra`, `local`)
`num_clips`	`int`	`5`	How many epic moments to detect
`top_k_ideas`	`int`	`3`	How many ideas to extract for remix
`mood`	`string`	from tone	Override mood (defaults to style tone)

Style Profile

The style profile controls the voice, visual aesthetic, and personality of the generated video. Only tone and persona are required.

{
  "tone": "sarcastic and irreverent",
  "persona": "tech skeptic who has seen it all",
  "hooks_style": "rhetorical questions",
  "visual_aesthetic": "warm film grain",
  "vocabulary": "casual gen-z slang",
  "catchphrases": ["here's the thing though", "let me cook"],
  "avoid": ["clickbait", "fear mongering"]
}

The style flows through every production step:

Script narration — persona and tone override the default “virality strategist” voice
Avatar — presenter_look derived from persona description
B-roll — visual_aesthetic embedded in every broll prompt and the continuity brief
Background music — mood derived from tone instead of generic “viral TikTok energy”

Output

{
  "output_path": "/tmp/fabric_effects.mp4",
  "epic_ideas": {
    "themes": ["Strategic Negotiation", "Power Dynamics"],
    "key_ideas": [
      {
        "idea": "A coach's history of leveraging job offers reveals a consistent pattern...",
        "talking_points": ["..."],
        "emotional_core": "surprise, intrigue",
        "controversy_angle": "Is this strategic leveraging brilliant or disloyal?"
      }
    ],
    "original_creator_style": "analytical and opinionated...",
    "remix_opportunities": ["Create a 'Playbook' series...", "Host a debate panel..."]
  },
  "script": { "full_narration": "...", "segments": [...] },
  "duration": 29.4
}

Video to Shorts

Workflow: video/video_to_shorts

Extract viral clips from a longer video (podcast, talk, lecture) and produce short-form content.

CLI
REST

fabric run video/video_to_shorts \
  --input url="https://youtube.com/watch?v=..."

curl -X POST "$FABRIC_URL/v1/workflows/run?name=video/video-to-shorts" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "url": "https://youtube.com/watch?v=..."
    }
  }'

Pipeline

transcription → detect_viral_moments → extract_clips →
reframe_to_vertical → burn_subtitles → compose_final

YouTube Studio

Workflow: youtube/studio

Complete YouTube metadata preparation: transcribe, generate title options, create thumbnails, and write description with chapters.

CLI
REST

fabric run youtube/studio --input url="https://youtube.com/watch?v=..."

curl -X POST "$FABRIC_URL/v1/workflows/run?name=youtube/studio" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "url": "https://youtube.com/watch?v=..."
    }
  }'

Pipeline

transcription → title_generation → thumbnail_generation →
description_generation → collect_output

Output

{
  "title": "Selected title",
  "title_options": [...],
  "thumbnail": "/tmp/thumbnail.png",
  "description": "Full YouTube description with chapters...",
  "transcript": [...],
  "duration": 345.6
}

Topic to Platforms Chain

Workflow: composed/topic-to-platforms-chain

Sequential rethink chain — one topic produces 8 platform-native posts, each re-derived from the topic (not re-formatted from a single source). Counterpart to the parallel global/platform-adapt for cases where deeper platform adaptation matters more than throughput.

The order is deliberate: each step inherits constraints discovered by the prior one.

twitter → linkedin → instagram → tiktok → youtube → newsletter → threads → facebook

Twitter goes first because its 280-char limit forces the tightest articulation of the core idea; each later step expands or reframes from there.

CLI
REST

fabric run composed/topic-to-platforms-chain \
  --input topic="Why most prompt-engineering advice is wrong"

curl -X POST "$FABRIC_URL/v1/workflows/run?name=composed/topic-to-platforms-chain" \
  -H "Authorization: Bearer $FABRIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": {
      "topic": "Why most prompt-engineering advice is wrong",
      "platforms": ["twitter", "linkedin", "instagram", "tiktok", "youtube", "newsletter", "threads", "facebook"],
      "brand_voice": {
        "voice_description": "punchy, opinionated, anti-fluff",
        "target_audience": "solo founders building AI products",
        "niche": "AI tools",
        "platform_tone": {"linkedin": "narrative"}
      }
    }
  }'

Pipeline

rethink_chain  (one LLM call per platform, sequential, prior outputs as context)

Output

{
  "platform_posts": [
    {"platform": "twitter", "text": "...", "hashtags": [], "angle": "...", "char_count": 268},
    {"platform": "linkedin", "text": "...", "hashtags": ["..."], "angle": "...", "char_count": 1240}
  ]
}

Notes

Pass a BrandVoice object (or dict) as brand_voice to apply per-platform tone overrides; a bare string is also accepted for backward compatibility with global/platform-adapt.
Override or shorten the platforms list to skip channels.
Each step is its own LLM call — expect ~8× the latency and cost of global/platform-adapt. Use that workflow when you want fan-out, this one when you want platform-native depth.