| name | video-factory-create-flow |
| description | The main orchestrator for the Slalom Video Factory. Walks a user from "I need a video" to a ready-to-generate prompt package (and, on request, generated clips) via the Frame → Concept → Storyboard → Direct → Generate → Ship pipeline. Diagnoses intent, sequences the craft skills, and pauses for review at each handoff. Use this skill whenever someone wants a cinematic video, a brand film, a motion accent, or a background loop for a product website and hasn't already scoped it down to a single stage — "make me a hero video for X," "I need a background loop," "let's storyboard a launch film." For a single stage, route straight to that craft skill instead.
|
| type | skill |
| project | skills-library |
| plugin | video-factory |
| aliases | ["video-factory","video-create","video-pipeline"] |
| tags | ["type/skill","plugin/video-factory","scope/orchestrator","topic/pipeline"] |
| status | active |
Video Factory — Create Flow
Take a user from a video idea to a generate-ready prompt package, sequencing the four craft skills with a review pause at each handoff. Load ../../references/video-factory-foundations.md once at the start for the pipeline, registers, output types, and anti-clichés — don't re-explain them downstream.
When to invoke
- "I need a video / background loop / brand film for [product/page]" — full pipeline.
- The user has source material (a deck, a product page, a brand) and wants it turned into video.
- For a single stage only, skip this and call the craft skill directly:
video-factory-concept, -storyboard, -creative-direction, or -higgsfield-prompt.
Stage 0 — Frame (quick intake)
Don't hand the user a form. Confirm only what you can't infer, using AskUserQuestion for the discrete forks:
- Output type — background loop / cinematic film / motion accent / product-UGC ad. (Sets every downstream constraint.)
- Placement — where it plays (hero behind a headline? landing feature? social?).
- Runtime & aspect — length and 16:9 / 9:16 / 21:9.
- Brand — Slalom / Composable DXP / client custom.
- Feeling — the two-second emotional read.
Capture free-text (product truth, audience, reference clips) conversationally. Start 60_Digital_Manufacturing/Video_Factory/WIP/{run_name}/run_log.md and write 00_Brief.md. Per the user's rule, every run gets a WIP folder and a run_log.
Stage 1 — Concept
Hand to video-factory-concept. Produces 10_Concepts.md with 3–5 distinct directions and a recommendation.
Pause: present the concepts; user picks or remixes one before storyboarding.
Stage 2 — Storyboard
Hand the chosen concept to video-factory-storyboard. Produces 20_Storyboard.md — shot list with timing, camera, transitions, and a keyframe prompt per shot; safe-area and loop-seam noted for background loops.
Pause: confirm the shot plan reads right.
Stage 3 — Direct
Hand to video-factory-creative-direction. Produces 30_Direction.md — the locked look (register, pacing, camera language, light, grade, grain, brand, NOT list) plus the four checks.
Pause: sign off the look. Nothing generates before this sign-off — credits are real.
Stage 4 — Generate (prompt package, then optionally clips)
Hand to video-factory-higgsfield-prompt. Produces 40_Prompts.md — per shot: keyframe prompt + image model, motion prompt, exact video model + parameters, negatives. If the user wants clips now: generate keyframes first, review against the direction spec, then generate clips via the Higgsfield MCP, logging credits.
Pause: review keyframes before spending video credits; review clips before Ship.
Stage 5 — Ship
Final cut, color, and web encode: .webm + .mp4 + poster frame, < ~3 MB for a hero loop, the <video muted autoplay loop playsinline> snippet, and a prefers-reduced-motion note. Promote the package and final assets to the engagement or 40_Library/. Hands off to the vercel / contentful skills for delivery on the Composable DXP stack.
Operating notes
- One review pause per stage — the user should feel prompted, not interrogated.
- The pipeline is also à la carte: if the user already has a storyboard, jump to Direct; if they have direction, jump to the prompt package.
- Keyframe-first is non-negotiable (the models are image-to-video). Direction sign-off gates all generation.
- Default models: seedance 2.0 / kling 3.0 for video, nano banana pro / chatgpt image 2 for keyframes.
