Video Style — Edho Ferdian Mode

Two phases with one interface between them — the style pack:

1. Distill — measure reference footage into a pack. references/distillation.md.
2. Apply — generate or collect footage, grade and cut it against the pack, verify, hand off. references/application.md.

The principle everything else follows

Prompting cannot deliver a grade; measurement can. In practice, escalating colour instructions to a video model barely move the result toward a measured target, while a deterministic grade computed from the references lands on it in one pass at no generation cost. So:

• the generative model supplies content, motion, framing and lighting structure;
• the pack supplies colour and rhythm.

Generation prompts therefore contain no colour language at all — say "Colour: none. Render neutral. Grading is applied afterwards." — and keep what happens (the brief) separate from how it looks (the style steer), or style words turn into objects ("teal" becomes a teal prop).

Non-negotiables

• Paid provider calls are opt-in per run. Every stage must have a dry run that exercises plans, prompts, track layout and manifests without credentials or spending. A live run needs the user's explicit go-ahead and its own opt-in switch, not just an API key in the environment.
• An ambiguous provider timeout is not retried as a new paid job. Check the provider's request log first.
• Verify numerically before claiming the look matches — distribution shape, not only averages (application.md §6). A good-looking MAE has shipped visibly broken frames.
• One genre, one pack. Never silently reuse one look's measurements for another. A measured zero is data; distinguish it from a missing field.
• Rights. Reference footage and anything lifted from it (stills, plates) may be someone else's work; confirm the user may use it before minting assets, and never ship another app's UI or someone's burnt-in titles.
• A render is not a saved project or creative approval. Keep source assets and versioned project checkpoints; report what was actually saved.

Tooling

No code ships with this skill; the references specify the method precisely enough to implement per project. Typical stack: ffmpeg (decode, cut, composite, re-encode), OpenCV + NumPy (Lab statistics, masks, plates), PySceneDetect (shot boundaries), a LUT writer (.cube), optionally Blender (3D props) and a VLM for the text spec. Resolve current library and provider APIs live before writing against them — model IDs, endpoints and prices drift monthly.

References

• references/distillation.md — pack layout, the measurements that matter (zone chroma, median/MAD, contrast, background share), interface masking, adaptive cadence detection, overlay plates, grounding a VLM, LUT baking pitfalls.
• references/application.md — generating takes not shots, cutting rules, grading rules, overlays as elements, removing borrowed UI, the 3D branch, numeric verification, and NLE/Blender handoff.

External docs (fixed — see skill-authoring-edho-ferdian's canonical contract)

Before calling a video/image/3D generation provider, PySceneDetect, or the Resolve scripting API, resolve its current interface live (Context7 or the provider's model page) — never from memory. Full contract: skill-authoring-edho-ferdian §9.

Surgical changes (fixed — see skill-authoring-edho-ferdian's canonical contract)

When adjusting an existing pipeline, change only the stage the task names; do not re-measure a cached pack or re-render untouched shots as a side effect. Full contract: skill-authoring-edho-ferdian §10.

Language routing (fixed — see skill-authoring-edho-ferdian's canonical contract)

Communication to the user in Bahasa Indonesia; code, comments, prompts to generation models, and generated files in English — fixed, never ask. Full contract: skill-authoring-edho-ferdian §7.