A personal, full-stack
agentic video studio.

North Star

End-to-end production across every genre — product ad · cinematic · talking-head/avatar · sound & music · montage · explainer.

Not a model wrapper. A skill + orchestration layer that turns a short prompt + the real product into a finished, claim-safe, Hebrew-first ad.

The model is a swappable commodity. The genius is the orchestration, craft, critics, and cost — not any single model.

Positioning — the unoccupied triangle

The moat — consistency without distortion

The single hardest thing in AI video, and our crown discipline: the same thing never drifts.

• Product consistency — never let a model redraw the real product (i2v from the exact still; first+last keyframe lock).
• Character consistency — a built, named, ownable mascot that looks identical across every clip & episode → write ongoing stories about him.
• A brand that owns a consistent character owns a franchise, not a one-off ad.

The pipeline — one brief, ten orchestrated stages

1. Brief — Hebrew prompt + pick the real product + claim attestation
2. Prompt craft — video-prompt-improver fills MCSLA slots + a named camera move (iterative)
3. Still / Product-Lock — Flux still from the exact product, or a locked character ref
4. Composite (P4) — if product-in-scene: cutout + placement (on_table / in_hand)
5. Motion (i2v) — first+last keyframe lock; cheapest provider that clears the bar
6. Dual critics — Red Flag + Copy Director attack → revise → regenerate (the ping-pong)
7. Sound stage — VO + ambient + SFX + music duck → video_final
8. Montage — stitch multiple beats into one 15s piece
9. Finish — upscale · grade · captions (Hebrew RTL) · platform aspect
10. SHIP gate — claim-safe verdict, delivered as a link

Stage 2 · Prompt ping-pong

The prompt is drafted, attacked, sharpened

The skill doesn't take the prompt at face value. It fills a structured slot framework (subject · action · camera · lighting · atmosphere), then critiques its weakest slot and rewrites — logging the winning delta so prompts compound across every future job.

A named camera move (push-in, rack, dolly-out…) is mandatory — bias toward cinematic craft, never a safe default push-in.

Stages 3–5 · The lock chain

Still → Composite → Motion, identity pinned at every hop

Stage 6 · The ping-pong that guarantees quality

Adversarial dual-critic loop → SHIP

Two independent critics, research-based rubrics. They try to break the work — slop, distortion, weak hook, unsafe claim. The piece only ships when both are satisfied. The model never gets to redraw the product to "fix" a critique.

Stages 7–9 · Finishing

Sound, montage, and the edit layer

• Sound stack — edge-tts VO + ambient bed + SFX overlay + music ducking → video_final.mp4. Sound is a named Higgsfield gap we close.
• Montage — deterministic ffmpeg stitch of several beats → one 15s piece (no drift, measured durations).
• Finish — upscale · color grade · Hebrew RTL captions (ffmpeg ASS) · per-platform aspect presets.

Stage 10 · The human ping-pong

PRISM gates — Ben decides, agents execute

The studio is built inside a governed factory. The creative ping-pong with the human is deliberate and bounded:

• P6 gates (Ben only): Vision · Brand · Naming · Logo · Positioning — each delivered as a deployed link, not chat.
• Everything else autonomous — agents decide technical matters with documented reasoning.
• Prove-before-build — real output before more spec; over-build is flagged as a decision.

Capability rings — the build order

The cheap-but-premium engine

How we beat Higgsfield on price:

• Provider seam (ports.py) routes each job to the cheapest provider that clears the quality bar.
• Cheap by default, premium opt-in — WaveSpeed undercuts cost; Replicate stays for reach.
• Still-first, cost-efficient — storyboard → generate → orchestrate. Complexity stays hidden: you type a brief, the studio picks pipeline → provider → critics → finish.

Shipped & verified

• Ring 1 Crown — Product-Lock ad pipeline live (P1–P5). done
• Ring 4 Montage — 3 mixed-aspect clips → one 15.02s montage, render-verified.
• Ring 2 Sound — VO + ambient + SFX → video_final; live ben-family smoke ≈ $0.053.
• P7 Character — reference-image plumbing + cast refs; ref_fallback verified on a live flux-schnell run ($0.003).

Gate decisions — approved

• Ambition — approved in full: full-stack Higgsfield-class studio, all rings, product-ad crown.
• Providers — approved: add WaveSpeed alongside Replicate (cheapest-clears-bar routing).
• OpenMontage — approved: clean-room blueprint, no fork, AGPL-safe.
• First ring built — Ring 4 (montage), then R0 → R2 (sound) → R3 (talking-head) → R5/6.