← GPT-AI-VIDEO
DOCUMENT · 2.4 KB

docs/production-loop.md

Workspace snapshot · 09/04 14:52

AI Video production loop

Roles

RoleDefault ownerOutputGate
Evidence curatorCodexpublic source manifest, privacy class, exact claimsno unsupported claim
Story editorCodexsingle claim, hook, turning point, payoffthe sequence proves one claim
Frame designerCodex native image generation + Remotionfour-frame visual familyself score 20/24
Dissenting creative directorClaude -pindependent rejection reportno blocker, 20/24
Motion scoutlow-cost video model3–5 second draft variantsidentity, physics, prompt adherence
Finishing modelbest provider for the selected shotselected production clip80/100 video gate
Deterministic compositorRemotionexact Japanese, logo, UI, captions, CTAzero text/logo/UI corruption
Sound editordedicated TTS/BGM/SFX toolsnarration and mixno clipping, intelligibility
Learning recorderCodexcost, latency, attempts, failure class, reusable rulereproducible by another operator

Required loop

  1. Build a public-source manifest and a one-claim brief.
  2. Create hook, development, turning, and payoff frames as one family.
  3. Self-score every frame. Do not send failures to a paid video model.
  4. Run Claude -p against the brief, sources, and quality gate. Verify the reviewer's factual statements.
  5. Fix one defect per iteration. After the same generative defect twice, switch to deterministic composition.
  6. Compare one real PENGIN shot across up to three video models. Draft first, finish only the winner.
  7. Add exact Japanese, logo, website UI, and CTA in Remotion.
  8. Record actual cost, generation time, score, and failure reason after every run.

Approved budget envelope

The operator approved a total experiment budget of JPY 50,000 on 2026-08-30. External publication remains separately gated.

StageCap
real-asset provider comparisonJPY 5,000
production candidates and regenerationJPY 30,000
upscale and derivativesJPY 5,000
voiceJPY 3,500
music, SFX, mixJPY 3,500
provider failure and final reserveJPY 3,000

The first paid comparison will use the same approved keyframe and motion brief across the candidate models. A transaction still requires action-time confirmation in the payment UI, and uploads containing people or private data require explicit destination confirmation.