How we built the five-phase pipeline

The version that generated everything at once

The first prototype was a single long-running job. You typed a topic, the job called every model it needed — script, images, voiceover, layout, render — and 12 minutes later a finished MP4 dropped into a folder. It demoed beautifully. It was also unusable in practice, and the reason is worth stating plainly.

When a 12-minute job produces a bad result, you have no idea which decision was bad. The script misread the topic? An image is off-brand? The voice picked the wrong emphasis? You re-run the whole thing and pay for the whole thing again, hoping the dice land differently. Cost and latency both scale with the part of the pipeline you trust least.

Why five phases, not three and not ten

The phase boundaries are not arbitrary. Each one is drawn where the pipeline produces a durable, inspectable artifact — something you can look at, that can be retried on its own, and whose cost can be accounted for separately. That is what the split buys: observability, resumability and cost isolation.

Three phases would merge Planning and Assets — precisely the merge the monolith proved costly. Ten would slice at boundaries with no meaningful artifact between them. Five is the count where each phase produces something you can name, inspect and retry by itself.

A separate question is how many of those boundaries should stop and ask you. That answer turned out to be one.

Phase 1 is the only gate

If you only ever review one thing, review the storyboard. Phase 1 produces a complete scene-by-scene plan — narration text, visual intent, and pacing — and no pixels or audio have been generated yet. Editing a scene's script here costs a text re-prompt. Editing it after Phase 2 means regenerating that scene's image and voiceover. After Phase 4 it means a re-render.

So the gate after Phase 1 does the heaviest lifting in the whole system: it moves the most consequential corrections to the cheapest possible moment. Once you approve it, phases 2 through 5 run to completion on their own — assets, direction, render and publishing metadata, with nothing further to approve.

We shipped later gates and then removed them, for the reason in the section below: a gate you cannot act on cheaply is not a control, it is an interruption.

Your work is never lost between phases

A phased pipeline is only worth having if the phase boundaries are durable. Close the tab during Phase 3, lose your connection, step away for an hour, come back tomorrow — your production is exactly where you left it. Nothing has to be restarted, and nothing you already paid to generate is thrown away. What this means in practice:

• Walk away anytime. Closing the browser changes nothing — the run continues, and the state lives on the production rather than in your tab.
• A failure is recoverable, not lost. If a phase hits a snag, the production pauses there with a clear Retry — you never lose the run or the work before it.
• Hands-off when you want it. Trust a repeatable series? Auto-Approve answers the Phase 1 gate for you, so a validated workflow runs from input to publish-ready untouched.

Auto-Approve didn't weaken the gate — it proved it

A reasonable objection: if you let users turn the gate off with Auto-Approve, was it ever load-bearing? It was. Auto-Approve is for the run you've already validated — a weekly series on a brand preset you've watched succeed ten times. The gate still exists; you've simply pre-decided the answer. It is a Professional and Enterprise feature; on other plans the single Phase 1 gate always applies.

What the gate buys you on a new kind of production — a new topic, a new brand, a new format — is still there the moment you want it. Auto-Approve is a trust setting, not a pipeline shortcut.

Three things we tried that didn't work

One review screen at the end. Reviewers approved videos that were 90% right because re-running was the only alternative. The gate has to be where the cheap fix is, not where the finished artifact is.

Letting phases overlap to save time. Starting asset generation before the plan was approved meant approving the plan sometimes invalidated work already in flight. The phases are sequential on purpose; the parallelism lives inside a phase (all of Phase 2's assets generate concurrently), never across the gate.

A gate after every phase. This is the one we shipped and then walked back. Approving generated assets told people nothing the plan had not already told them, and by that point disagreeing meant paying to regenerate — so those gates collected clicks, not decisions. Four of the five were costing attention and buying nothing, so they went. The later approval controls still visible in the interface are vestigial.

Auto-retrying failed phases silently. A silent retry that also fails just burns tokens and time. A failed phase now surfaces immediately with the error and a manual Retry, so a misconfigured run gets fixed instead of looped.

Further reading

• How It Works — the five phases walked through end to end.
• One review gate, not five — the argument for the number, in full.
• Phase 2–5 tabs — what each phase tab shows once it completes.
• Auto-Approve Mode — when to pre-decide the gate.
• Multi-language fan-out, explained — how one approved plan becomes 24 videos.