# Seedance-2.0 harvest — evaluation & coverage map

What we took from the external **`Emily2040/seedance-2.0` "Seedance 2.0 Skill OS"**
(a fork lives at `iswhat/seedance-2.0`), and why. MIT-licensed, docs-only (28
skills + 56 references; its Python is validation scripts, not a transport).
Everything here is pinned to commit **`63b32dc`**.

## Decision: harvest, don't adopt (option A)

The repo is a **knowledge/methodology** layer — it cannot render a frame.
videoclaw already owns the hard parts (native-seedance transport, Asset Library
identity, content-filter, cost). So we **harvest its additive knowledge into
videoclaw's deterministic typed registers**, rather than adopting a parallel
Markdown skill system (which would fight the Track-A-inert/Track-B-typed design in
`seedance-skill-loader.ts` and the concierge front door). Each harvested piece is
a pure module + an advisory folded into an existing surface (or a small opt-in
flag / read-only command), default-safe, no spend.

## The 14 Core Pipeline skills → videoclaw coverage

| Seedance skill | videoclaw coverage | Status |
|---|---|---|
| `seedance-interview` / `-short` | concierge (Clawbot) front door + `studio` recipes | ✅ covered |
| `seedance-sequence` | `execute --auto-chain` (`runAutoChain`) + storyboard + studio | ✅ covered |
| `seedance-continuation` | `continuation-handoff` (chain-depth + failure-atlas) + `chain-from`/`reroll-scene` | ✅ **shipped** (#218) |
| `seedance-prompt` / `-short` | `filmmaking-prompts` + `multi-shot` (+ anti-slop for compression) | ✅ covered |
| `seedance-camera` | `cinematography.ts` + `shot-grammar.ts` (shot size / angle / movement) | ✅ covered |
| `seedance-motion` | `cinematography` movement grammar + FRAME MAP | ✅ covered |
| `seedance-lighting` | `cinematography` LIGHTING / LIGHTING_PROSE registers | ✅ covered |
| `seedance-characters` | Seedance Asset Library + `subjectLockBlock` + character-consistency | ✅ covered (strong) |
| `seedance-style` | `cinematography` genre/style + `seedance-content-filter` (IP-safe) + brand-definition | ✅ covered |
| `seedance-vfx` | **was a gap** → `vfx-register.ts` + `multi-shot --vfx` | ✅ **shipped** (#222) |
| `seedance-audio` | `audio-platform` + `diegeticAudioLine` + assemble mix-plan + `multi-shot --dialogue` | ✅ covered |
| `seedance-pipeline` | videoclaw **is** the pipeline | ✅ covered |
| `seedance-recipes` | `studio` recipes + `storyboard-templates` + GENRE_DEFAULTS | ✅ covered |
| `seedance-troubleshoot` | **was a gap** → `vclaw video diagnose` (output-diagnosis tree) | ✅ **shipped** (#223) |

**Conclusion:** of the 14 Core Pipeline skills, 12 were already well-covered by
existing videoclaw machinery. The two genuine net-new gaps — **VFX vocabulary**
and a **general output-troubleshoot tree** — are now shipped.

## Reference-doc harvest (beyond the 14 skills)

| Reference | videoclaw landing | Status |
|---|---|---|
| `anti-slop-lexicon` | `seedance-antislop.ts` → `prompt-lint slop` advisory | ✅ **merged** (#216) |
| `retake-protocol` | `retake-protocol.ts` → `reroll-scene` shot-log + 2 hard rules | ✅ **merged** (#217) |
| `failure-atlas` + chain-depth | `continuation-handoff.ts` → `AutoChainReport.advisories` | ✅ **merged** (#218) |
| `reference-transfer-contract` | `reference-transfer.ts` → `prompt-lint reference-transfer` | 🔵 open (#221) |
| `model-mechanics` | vendored as inert Track-A knowledge (this PR) | ✅ this PR |
| `allocation-model` | over-allocation lint (name primary/secondary/economized spend) | ⏳ backlog |

## Remaining backlog

- **allocation-model** — a `prompt-lint` advisory that warns when one packet
  demands the full budget at once (a perfect face + bold motion + a crowded scene
  + readable text). Deferred: it shares `prompt-lint.ts` with the open #221, and
  its "demands everything" detection is heuristic — worth a careful, conservative
  pass once #221 lands.
- Deeper continuation surgery (completed/reserved **beat exclusion**,
  observed-state-overrides-planned prompt rewriting) — needs spend-path changes;
  intentionally out of scope so far.

## What we deliberately did NOT take

- The 28-skill Markdown OS as a parallel system (fights the typed-register design
  + the concierge front door).
- The repo's dated per-surface **API-status / pricing** claims — videoclaw has
  fresher ground truth; those go stale.
- The multilingual (zh/ja/ko/es/ru) vocab — out of scope for now.

## Attribution

All harvested content is MIT © 2026 Iamemily2050
(`github.com/Emily2040/seedance-2.0`, pinned `63b32dc`), adapted into videoclaw's
typed registers. Each module/PR carries its own attribution line.