The production playbook

1. 01Intake

   One product photo + minimum context

2. 02BriefHuman

   One structured brief drives every deliverable

3. 03Route

   Each asset to its cheapest sufficient model

4. 04Generate

   Stills, adaptations, motion, audio

5. 05GateHuman

   Fidelity · text · brand safety

6. 06Deliver

   Named, versioned, ready to ship

Two human checkpoints, deliberately placed: at the brief (before any spend) and at the gate (before delivery). Everything between is automated.

BRIEF          mood · setting · palette · headline EN/FR
     +
FORMAT BLOCK   "Square 1:1 promo tile. Render "{headlineFR}"
                in clean bold type, high contrast…"
     +
STYLE SUFFIX   "commercial retail quality, sharp focus"
     =
  PROMPT  ────────────────▶  nano-banana-pro

Each deliverable has a prompt builder, not a prompt. Change one field in the brief and eight prompts update. That is what makes versioning — EN/FR, seasonal, per-format — a parameter instead of a project.

• Subject and action first; camera, light and setting second; style descriptor last.
• Real product attributes injected from vision analysis — never invented.
• Video prompts carry an explicit Audio: cue; native-audio models reward sound design written into the prompt.
• Negative prompts are maintained per style, with artifact patterns added as QA finds them.
• No brand names or logos in generation prompts — lockups composite in post, where they're controlled.
• For product motion, references beat a first frame: reference-to-video models hold identity across the take, addressed positionally as [Image1], [Video1] with a stated job each.

The live table lives on the model landscape page and is re-evaluated monthly.

Three lines that do not move

On a real product, where the line actually falls

This is the most specific rule in the set, so it gets the most specific treatment. Everything on the left is production craft. Everything on the right changes what the customer thinks they are buying.

Depicting people

What has to survive an audit

Approved tools only

Platforms cleared under the applicable technology, security, procurement and privacy requirements. That is a list someone maintains, not a judgement call at 6pm.

What never gets uploaded

Confidential information, personal information, biometric source material, talent recordings, licensed content, or third-party assets whose AI-processing rights have not been confirmed — unless that specific system is authorised for it.

Retain the real source

Where AI materially assists final product imagery, the authentic source frame stays in the asset-management system and stays traceable to the delivered creative. If you cannot produce the original, you cannot defend the final.

The record

Platform or vendor, what the AI component actually was, the approved source asset, confirmation of commercial-use rights, talent consent where relevant, and any specific exception granted. Meaningful provenance — not a log of every routine retouch.

Accountability does not transfer

Marketing, agency and production partners own the finished asset whether or not AI touched it. There is no version of this where the model is responsible.

Two questions that settle most of it

Where this is enforced rather than promised

A guideline nothing implements is a hope. These are the places the studio in this demo makes the rule structural — so following it is the default path, not the disciplined one.

Prices are never invented

Vision autofill returns an empty field rather than a guess when no price is legible on the pack, and the UI flags it as needing a human. A wrong price is a compliance failure, not a typo.

Reconstructed angles are labelled

Packshot views the model extrapolated rather than grounded in a supplied reference are marked for label QA, so nobody mistakes a plausible back-of-pack for a photographed one.

The prompt ships with the asset

Every generated output exposes the prompt that made it, which is the provenance record the guidelines ask for, produced as a by-product rather than as homework.

Product identity is pinned, not hoped for

Reference-to-video binds the pack to a supplied still and instructs against drift, because a product that morphs mid-shot fails the fidelity test even when nobody intended it to.

The demo states its own limits

Where the pipeline cannot guarantee something — music that is not frame-synced, a take that needs an alignment pass — it says so in the interface rather than in a footnote.

Two things that are not exceptions

This page expires. These guidelines get reviewed against regulation, platform and broadcaster requirements, talent agreements, campaign learnings and the quality of synthetic media itself. The clause most likely to move first is the broadcast one — synthetic on-camera talent is a conservative default about consumer and legal acceptance, not a permanent technical judgement.

Every generated asset in the demo exposes its prompt — the pipeline shows its work. That's the teaching model: every workflow visible, every decision documented, every playbook written so the second person, and the tenth, can run it.

Capability that lives in one person's head isn't a studio. It's a bottleneck with a title.