Prompt → Visual Asset QA before anything ships

1. Why this recipe exists

The I/O 2026 generative-asset story is that generation is cheap. The expensive part is the QA grid that stops the wrong variant from shipping. Most teams have a sketch of this loop in someone's head. This recipe writes it down.

2. Lock the prompt

The recipe ships with one sample prompt: compact dashboard hero, dark slate, one chart, no text. Locking the prompt means it does not change between variants. If it changes, you are running a different experiment.

3. Load four placeholder variants

No live model call. The recipe ships four pre-rendered placeholder cards so the loop is testable without a paid generation pass. Each card stands in for what a real Imagen / Veo output would look like, including the failure modes.

4. Run the QA grid

Four checks per variant:

• Brand safety. Anything that violates the brand bar fails.
• On-prompt fidelity. Off-theme or off-style fails.
• Embedded text. Any stray rendered text fails. Public surfaces should never ship model-rendered text.
• Alt-text candidate. A one-line alt-text the variant could ship with.

5. Pick one + log the rest

Exactly one variant ships. The other three land in a rejected-asset log with a one-line reason. The log is the durable artifact, so six weeks later you can re-read why a variant was passed over.

What the artifact looks like

The hero renders the sample QA grid for the bundled prompt: variant A ships, variants B / C / D land in the rejected log with one-line reasons (stray text, off-prompt, busier-than-needed). Copy the artifact text and paste it into the asset PR description.

Public-safe rails

• No live generation. Recipe uses pre-rendered placeholder cards so the loop runs $0.
• Embedded-text check. Any variant with stray rendered text fails QA.
• Brand-safety check. Runs on every variant before the row is marked shippable.