benchrThe playbook
A brief to a finished commercial
Nine stages, and the expensive one is not the one people expect. Generation is cheap; continuity and selection are the work.
9 stages Two to five days for thirty seconds, most of it discarding takes Gemini API · OpenAI API · claude.ai and the Claude apps Complete
Every stage names what goes in, what to do, why it sits here and what comes out, and the chain says where it breaks.
benchr has not run these capabilities itself. Each stage points at a record describing what a provider documents, read on the date shown.
- 01
Idea
Your work- You start with
- A brief, a product and one thing you want the viewer to feel
- Why this stage
- Everything downstream inherits this. A vague idea produces expensive vagueness forty generations later.
- Do this
- Write the single sentence the ad has to land. If you cannot, do not start generating — the model will happily make something beautiful about nothing.
- Or
- Skip only if the client handed you a locked concept.
- You end with
- One sentence and a product
- 02
Script
Your work- You start with
- The sentence and the runtime
- Why this stage
- Runtime is a hard constraint that generation will not respect for you.
- Do this
- Write to time, out loud, with a stopwatch. Thirty seconds is roughly seventy-five words of voiceover, less if anything happens on screen.
- Reach for
- Any capable chat model; this is the one stage where the model is a writing partner rather than a renderer.
- You end with
- A timed script
- You start with
- The real product
- Why this stage
- This is the stage that decides whether the finished piece looks like one film or forty samples. Do it before anything moves.
- Do this
- Photograph or fix the product once and treat that frame as the only description of it that exists. Build a small reference sheet — three angles — and never describe the product in words again.
- You end with
- A locked reference sheet
What goes wrong here
- 04
Shot list
Your work- You start with
- The script and the reference sheet
- Why this stage
- A shot list is where you find out the ad is too ambitious, while that still costs nothing.
- Do this
- One line per shot: size, lens, movement, what must not change. Mark which shots show the product fully — those are the ones you will re-run most.
- You end with
- A numbered shot list
- You start with
- Shot list plus reference sheet
- Why this stage
- Approving a still costs one generation. Approving a clip costs many. Get the frame right while it is cheap.
- Do this
- Generate the key frame for each shot from the reference, changing one variable at a time. Approve frames, not vibes.
- Or
- If the shot has no product in it, you can generate straight to motion and save a step.
- You end with
- One approved still per shot
- You start with
- Approved stills
- Why this stage
- Image-to-video inherits the frame you approved. Text-to-video starts a new argument about what the product looks like.
- Do this
- Animate each approved still separately — one camera move, one action per clip. Expect to discard most takes; that is the job, not a failure.
- You end with
- Short clips you can cut
What goes wrong here
- You start with
- The timed script
- Why this stage
- Recording once against a locked script is faster than regenerating narration to fit a cut that keeps moving.
- Do this
- Record the narration against the shot list, not against the edit. Keep the takes; you will need alternates when the cut changes.
- You end with
- A narration track and alternates
- 08
Sound
Your work- You start with
- The cut and the narration
- Why this stage
- Sound is where a rough AI cut starts reading as a finished piece, and it is the cheapest quality you can buy at this stage.
- Do this
- Music and effects last, cut to the edit. Licensed library music is usually the right call over generated music for anything a client will run.
- You end with
- A mixed track
- 09
Edit
Your work- You start with
- Clips, narration, music
- Why this stage
- The edit is where continuity failures either get hidden or become the thing everyone notices.
- Do this
- Cut on movement. Put the hardest continuity across a cut rather than inside a shot — the same trick a live-action editor uses, for the same reason.
- You end with
- The finished commercial
What goes wrong here