Model guide.

Prompt and script plan

A prompt is first checked and converted into a script plan. The plan keeps the user prompt, chosen tool, language, duration, model route, visual scene count and chunk plan together before any provider output is claimed.

If a prompt is too vague, blocked by safety rules or rejected before a provider stage starts, the job should return a clear error and no generated asset should be shown.

Script-plan failure: keep the prompt and settings, then regenerate the plan after the user edits the prompt or source notes.

Image failure: keep the script plan and retry the image stage without rebuilding the whole job.

One scene can still need several video chunks

Baby Podcast, Baby TED Talk and Baby Myth-Buster use one stable visual scene: the same character and set stay consistent. That is different from provider video chunks. A 30-second Baby Podcast default with Veo uses 4 fixed 8-second chunks, then the final output is stitched and trimmed to the target length.

Veo routes should stay locked to 8-second chunks. Users should not see arbitrary shorter or longer Veo scene seconds. Seedance and Kling-style routes are handled differently: their estimate can move up or down when the selected output seconds change.

Model and route readiness

A tool can show settings and estimates before paid generation is enabled. Generation should only open when the selected image, video, voice, lip-sync, stitch and cleanup paths are generation-ready.

Image, video, voice and lip-sync are separate checks. A better video model does not remove the need to check voice and lip-sync behavior. Avatar-style backup paths are not the primary plan for this workflow; primary tests should use current video models plus voice/lip-sync where possible.

Voice, audio and lip-sync

Voice is its own stage, with tool-specific voice defaults and optional voice selection when the tool supports it.

Lip-sync is also separate. The primary path should preserve finished voice and video assets where possible. If lip-sync fails, retry lip-sync only instead of rebuilding the whole job.

Generation moderation or rejection means the service stopped before usable output. The retry should route the user to prompt or asset edits, not silently charge again.

Credits, settlement and refunds

The tool shows a credit estimate before generation. The current MVP deducts the estimate at job creation and tracks every stage separately after that: estimated, deducted at start, settled, refunded and held for support.

A failed or cancelled chargeable stage can be refunded with a stage-specific ledger reference. If actual provider cost is lower than the estimate, the job can settle lower. If actual cost would be higher, the MVP should not auto-charge above the estimate without explicit confirmation or support review.

Prompt JSON and JSON Stitch

Prompt JSON shows the structured plan behind the simple prompt view. It should expose scene, timing, model and caption instructions, but never provider credentials, secrets or internal tokens.

JSON Stitch is planned for owned generated clips, approved public recipes and whitelisted internal demo fixtures. Until runtime validation is complete, treat it as coming soon: no arbitrary uploads, external URLs or local file paths should be accepted for final export.

Metadata cleanup

Export cleanup can normalize container metadata and remove C2PA/provenance metadata where the file format and evidence check allow it. It does not promise SynthID removal, pixel watermark removal, audio watermark removal or guaranteed platform AI-label bypass.

If cleanup fails after a final video exists, the finished render should stay attached and the cleanup stage can retry or be held for support with a visible warning.

Running jobs, email and multiple generations

A job keeps its own stage history, output URL and email notification preference. Users can close the page and return to the job detail page while processing continues. Multiple jobs can run independently as long as credit and route checks pass for each job.

If an email notification fails, the job page remains the source of truth for status and output links.

Sources and verdict labels

Baby Myth-Buster can accept optional source or context notes. Without a configured fact-check provider, verified verdict mode stays unavailable and output must be labelled as an entertainment explainer rather than verified advice.

Verdict labels should distinguish likely true, likely false, mixed, needs sources and manual review.

Medical treatment, legal advice, financial advice, dangerous activities, child-safety claims, elections/current political claims and targeted personal allegations require manual review before paid generation.