Claude Skills → Seedance 2.0
The video is not really about Seedance. It's about refusing to use one prompt formula for four different kinds of content — and instead loading a different instruction file per scene type. Seedance is just the renderer at the end.
One formula, four content types, average results
A cinematic action scene, an animated sequence, a product ad and a dialogue scene each need different camera language, different pacing and different reference logic. Forcing the same prompt skeleton across all four is, in his words, exactly why most results still look average.
The fix: one Claude Skill per scene type — each built around the prompting patterns that actually work for that kind of shot.
What a Claude Skill actually is
- A plain markdown file sitting in a
skills/folder. - Contents: the skill name, a set of instructions telling Claude how to think about one task, and a couple of examples showing the input → output pattern you want.
- On load, Claude stops behaving like a general assistant and follows that file's rules for everything it produces afterward.
- He shows the file on screen (01:26) and calls the structure "simple, but incredibly effective" — no tooling, no API, no code.
The load-bearing idea: a skill is an opinion about process, not a magic prompt. It changes what Claude asks for and what shape it outputs — before any words hit the model.
Why one formula fails
Camera movement and shot planning across cuts.
Clear style, simple physics, particles/blur on fast movement.
Clean framing, product focus, bright commercial lighting.
Lip sync, emotion, shots built around what's being said.
1 · Cinematic film
Owns the camera language people leave out: dolly moves, crane shots, rack focus, and shot discipline across multiple cuts. That's his stated difference between a generic clip and one that feels pulled from a real film.
Test 1 — brief: "create a 15-second shot of a medieval war between two armies." Answers: dark fantasy / brutal / 3 shots / mixed camera energy. Output:
Test 2 — a character reference sheet uploaded alongside the concept ("gritty ultra-realistic one-take classroom fight"). Answers: stylized martial arts, The Raid as reference, cold and detached. Constraint carried into every shot: handheld operator following throughout, CHARACTER STAYS CONSISTENT ACROSS EVERY SHOT.
2 · AI animation
Deliberately inverted priority: style consistency and physics description over realism or camera language. Reuse the cinematic structure here and he says the output goes flat, because the model has no idea what visual language you're in.
Test 1 — "cinematic high school fight in the rain." Answers: Shonen kinetic (Tokyo Revengers / Wind Breaker as reference), stylized grounded physics, two delinquent students, concrete alley behind the school. Structure worth copying: the entire style block lands at the very top — bold outlines, strong shadows, gray rainy palette — before any shot is written. Micro-detail carried into shots: water splashing from footsteps, rain falling, small flashes on fist impact.
Test 2 — Ghibli farm scene, woman picking tomatoes in soft morning light. Same skill, different style. What made it stick: literal technical parameters written into the prompt — watercolor backgrounds, gentle line work, 12 fps on 2-second character animation.
3 · Product ad
Clean hero framing, commercial lighting, product front and center in every shot.
Test 1 — ice cream brand ad from a product reference image. Answers: 3/4 turntable, sunlit rooftop lifestyle, mid-paced commercial cuts, 16:9.
People are supporting cast; the product is the hero in all three. Test 2 — UGC serum ad, swapped to eye-level held-to-camera, sunlit bedroom, "creator rhythm," 9:16 for TikTok/Reels, full lip-sync dialogue written into every shot, ending on a close-up framed for a CTA.
4 · Dialogue scenes
Where he says most people fall apart: lip-sync cues, emotional direction, shot structure and audio detail all have to work together, or you get robotic delivery and drifting sync.
Test 1 — police interrogation with a fourth-wall break. Answers: tense and short, simple over-the-shoulder framing, a sharp pause at the break, English.
Test 2 — before/after confidence transformation at the same desk. Key technique: camera framing stays identical across both emotional states so the performance is the only variable. States get described physically — slouched posture and shoulder tension, then open chest, steady breathing, quiet confidence. No fourth-wall break until the final shot.
Four mechanics worth stealing
| Mechanic | What it does / why it matters |
|---|---|
| Intake gate | The skill refuses to write a prompt until it has pulled four answers out of you. This is the product — not the prompt text. It converts vague intent into decided parameters. |
| Per-shot schema | Every output: subject, action, environment, camera, style, constraints — broken down per shot with timestamps. Fixed schema beats free prose for anything with cuts. |
| Reference transcription | Upload a character sheet or product photo and the skill writes the description from the image instead of from your prose. That's what held character and product consistency across shots. |
| Duration-scaled length | Prompt density is bound to runtime (15s → 3 shots). Plus the contrast trick: hold framing constant so only the variable you're testing changes. |
Building the skill
- Claude → Add files → Skills → Skill creator → paste your meta-prompt.
- That meta-prompt has two jobs: (a) always collect genre, tone, shot count, camera energy and duration before generating anything; (b) always output a full structured Seedance prompt — subject, action, environment, camera, style, constraints — broken down per shot with timestamps.
- Hit generate, answer the few setup questions, and the skill is ready in seconds.
- His closing pitch: this is the way to get consistent output without starting from scratch every time. Free prompt pack is linked in the description (also linked in the sidebar here).
Caveats
- No A/B test. Nothing is compared against a generic prompt on the same concept, no seeds, no repeats. Every "amazing" is a hand-picked single take.
- It's an ad. Higgsfield referral link in the description, and the render setup is their UI (1080p / 15s / 16:9). The technique survives without them; the workflow shown is theirs.
- One verbal slip: at 06:22 he says "paste it into SeaArt" — everything shown is Seedance.
- The skill files are not shown in full. You see the folder structure and one file's shape; the actual instruction text has to come from the prompt pack or be rebuilt.
For FightFlix
The portable piece isn't Seedance — it's the intake gate + fixed per-shot schema + reference transcription. That maps cleanly onto prompt generation for film-03: hold a four-question gate per scene type, emit timestamped shot blocks with subject/action/environment/camera/style/constraints, and transcribe character descriptions from uploaded sheets rather than rewriting them each episode. Zero pipeline changes; it tightens shot-level consistency at the prompt layer.
Next step if useful
Pull the author's prompt pack doc, rebuild the four skills as FightFlix-ready skill files with your own character/episode schema folded in.