Feed it the angle
Generate the statics
Render the video
Build the variant matrix
Review and release
Context & problem
Research tells you what to test. Then production makes you wait for it. A single video ad means a script, a shoot or an editor, captions, sound, and two weeks of back and forth, and by then the angle is stale or the budget moved on.
The pattern I kept seeing in DTC and agency work: the ideas were never the bottleneck, the variants were. Testing five hooks against two formats means ten assets, and nobody has a production team sized for that per week.
So I built production as a pipeline. An angle or a script goes in, finished creative comes out, and the human hours go into deciding what runs instead of assembling files.
How it works
The diagram above shows the flow; here is what each step does. Deterministic stays deterministic, and an agent only shows up where judgement, language, or synthesis is actually needed.
- 01Deterministic
Feed it the angle
Input is a brief, an angle, or a script, plus the brand's assets. Structured intake, so the same input always produces comparable output.
- 02Agent
Generate the statics
Image ads get generated in the brand's look, with hook and copy variants per angle. Generative work that used to cost a designer per variant.
- 03Agent
Render the video
A script becomes a finished video sales letter: talking AI avatar, matched B-roll, captions, and sound mix, rendered end-to-end without an editor.
- 04Deterministic
Build the variant matrix
Hooks, formats, and aspect ratios get combined into a test-ready set. This is combinatorics, not judgement, so it stays rules-based.
- 05Deterministic
Review and release
Every asset lands in a review set before it touches an ad account. A person picks what runs.
Checkpoints & logging
The gate sits between production and the ad account. Creative going out under a brand is not the place for unreviewed output, so the pipeline produces and a person releases.
Every asset is traceable to its input: which angle, which script version, which brand assets. When a variant wins, you can reproduce why. When one looks wrong, you can see where it went wrong.
Stack — and why
Orchestration runs on n8n so every production step is inspectable. Image generation and the VSL pipeline use generative models per asset because that is the actual production work, while intake, the variant matrix, and dispatch stay deterministic so a batch of ten variants behaves like a batch, not ten surprises. The avatar, captioning, and sound steps are wired as swappable stages, because that corner of the stack changes fast.
Results
The pipeline is shipped and has produced finished VSLs end-to-end, script to rendered video, without an editor in the loop.
It powered the creative for an outbound and ad motion across 10,000+ B2B contacts, my own system, so I saw the production economics first-hand: variants stopped being the constraint.
Creative testing stuck on production?
If you already run creative research, this engine is the natural next stage. The audit shows where it would earn its keep in your setup, in 30 minutes.