Production pipeline (Steps 1–15)

Every production runs through a fixed sequence of pipeline steps, each fulfilled by an AI model. You don’t run these by hand — they execute automatically once you start a production — but knowing what each does helps you read progress and troubleshoot.

🎬 How the pipeline works

Prefer to read? Open the step-by-step transcript

The pipeline turns your brief into a finished deliverable in five phases:

1. Write (Steps 1–4) — brief → narrative → beats → per-scene script.
2. Visuals (Steps 5–7) — source/generate images, safety-check them, pick the best.
3. Audio (Steps 8–9) — voiceover + soundtrack.
4. Video & variants (Steps 10–12) — compose video, make per-platform variants, write SEO.
5. Package (Steps 13–15) — attribution, final deliverables + transcript, gate checks.

Watch each step’s status badge advance on the production detail page.

The five phases

Phase	Steps	What happens
Write	1–4	The brief, narrative, beat outlines, and per-scene script (voiceover + image prompts + cues) are written.
Visuals	5–7	Images are sourced (stock), generated (AI), or operator-supplied; each is safety/rights-checked; the best per scene is chosen.
Audio	8–9	Voiceover is produced per scene (TTS), and a soundtrack/ambience is added.
Video & variants	10–12	Scenes are composed into video, per-platform aspect-ratio/duration variants are made, and SEO metadata is written.
Package	13–15	Attribution (licences/credits) is aggregated, final per-platform deliverables (with captions from Step 8a) are packaged, and a final gate runs before distribution.

Step reference

Step	Name	What it does
1	ProductionBriefGenerator	Picks subject + duration + structure
2	NarrativeGenerator	Generates per-Act narrative
3	SequenceBeatGenerator	Per-beat outlines
4	SceneScriptGenerator	Per-scene VO + image-prompt + FX cues
4a / 4b	ImageAcquisitionRouter / VideoAcquisitionRouter	Route each scene’s image (AI / stock / operator) and video to a source
5 / 5b / 5c	Image acquisition (stock + AI + operator-supplied)	Sources images per scene
6	ImageAnalyser	Safety + copyright + commercial-use evaluation
6a	AudioDescriptionWriter	Audio-description scripts for visually-impaired audiences
7	ImageProducer	Picks the winning image per scene
8	VoiceoverProducer	Per-scene TTS (ElevenLabs / OpenAI / Polly / Azure / …)
8a	CaptionGenerator	Closed captions (VTT / SRT)
8b	AudioAssembler	One mixed audio track, for audio-only output
8c	TranslationDubbingProducer	Translated, re-voiced voiceover in other languages
9	SoundtrackProducer	Background music + ambience
10 / 10b–d	Video composition (Shotstack / Creatomate / Cloudinary / Mux / Talks / Avatar)	Per-scene videos
11	PlatformVariantProducer	Per-platform aspect-ratio + duration variants
12	SEOGenerator	Per-platform titles / descriptions / hashtags
12b	StockKeywordOptimiser	Stock-marketplace keywords and category codes
13	AttributionPackager	Licence + credits + rights aggregation
14	DeliverablePackager	Final per-platform deliverables
14a	TranscriptGenerator	Searchable text transcript of the voiceover
15	GateEvaluator	Gate checks before distribution dispatch

Note

The exact steps that run depend on the account’s active steps — a podcast-style account skips video composition, for instance. Every production copies the account’s list when it starts, and you can adjust it for one production in Active steps on the New Production form. The pipeline activates only the steps on that list.

Reading step status

On the production detail page each step shows a badge:

• Queued — waiting for an upstream step.
• InProgress — running now (pulsing).
• Complete ✅ — done.
• Failed ❌ — permanent failure. Steps 8, 10, 12 and 15 can be retried from the Console; any other failed step is moved past with Manual Override — see Troubleshooting.
• AwaitingRelease — paused for an operator to release it (Controlled/Step-through mode). Not a failure — this is the pipeline waiting on a person, not stuck.
• ManualOverride — someone resolved the step by hand with Manual Override and the pipeline moved on.

The step verbs — Release Step / Reject, Retry Step, Refine & Re-run (Steps 8, 10 and 12) and Manual Override — are available to the account’s CustomerAdministrator, an assigned AccountOperator and platform admins. See Controlling a running production.

Rough timing: Write ~1–2 min, Visuals ~3–5 min, Audio/Video ~5–10 min depending on length, Package ~1–2 min — on top of however long a held step waits AwaitingRelease.

Quality & gates

Each AI-generating step records a quality score; a background check compares it against your account’s quality floor and flags regressions. The final GateEvaluator (Step 15) runs the checks that must pass before a production advances to distribution.

Tip

If a production stops at AwaitingRelease, it’s waiting on an operator to release it from the production detail page — that’s normal under Controlled/Step-through mode, not an error. If it stops at Failed, read the step’s error and either retry it (Steps 8/10/12/15 only) or use Manual Override. See Controlling a running production and Troubleshooting.