菜单
Use when writing or improving the creative content of a venice-video-mcp production
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when calling any of the venice-video-mcp tools and you need a concrete, copyable example of the tool arguments. Provides one example per action across all 6 tools, with every argument spelled out and reasonable defaults explained.
Use when the user asks the agent to drive the venice-video-mcp server end-to-end
Use when a venice-video-mcp tool call fails, returns ok=false, hangs, produces wrong output (wrong aspect ratio, character drift, duration error, missing audio), or when the agent needs to know which production gotchas to avoid before driving an expensive operation. Catalogs every known failure mode with cause and fix.
| name | venice-mcp-directing |
| description | Use when writing or improving the creative content of a venice-video-mcp production --- the episode concept, per-shot prompts, dialogue delivery, retakes, or multi-episode continuity. Bridges the separately-installed Seedance 2.0 Skill OS (the "director's brain") onto the venice-video-mcp tool surface so shots are DIRECTED (one intention drives camera, light, blocking, performance, sound) instead of decorated with "cinematic" adjectives. Load it alongside venice-mcp-pipeline. |
The pipeline skill tells you which tool to call. This skill tells you what the creative content should say before you call it.
It is a thin bridge to the Seedance 2.0 Skill OS (emily2040/seedance-2.0) --- a separately-installed agent-skill package that is pure directing/prompting knowledge with zero execution code. Venice ships Seedance 2.0 (+ Fast) as a video model family, and the harness routes most character work through Seedance-family R2V, so that directing knowledge applies almost verbatim --- once you subtract the parts the venice stack already owns.
Mental model:
Seedance OS is installed separately (it is not vendored into this repo). Typical targets: .cursor/skills/seedance-20/, ~/.claude/skills/seedance-20/, or .agents/skills/seedance-20/. See the venice-video-mcp README "Directing layer" section.
Direct the scene, don't decorate it. Most prompts ask a model for a "cinematic look" and stack adjectives. A director instead asks what the scene is doing --- the turn, the point of view, the power, the subtext --- names one intention, and derives camera, lens, light, blocking, performance, and sound from that single intention. Then holds one directorial voice across every shot of the episode and across episodes.
epic cinematic shot of a woman reading a letter, emotional, beautiful lightingMedium close-up, eye-level; she lowers the letter and her hands go still as a slow push-in arrives; soft window light keeps her face plain; near-silence with one chair scrape --- the realization lands in the stilled hands, not a word.A reveal is not lit, framed, blocked, or performed like a goodbye. Refuse the generic answer; derive the specific one. This is exactly what belongs in a shot's description and in the episode.workshop concept.
| When you are... | Load (Seedance OS) | Compile its output into (venice-video-mcp) |
|---|---|---|
| Turning a vague idea into a brief | seedance-interview / seedance-interview-short | the episode.workshop concept string |
| Planning a longer / multi-clip story | seedance-sequence | the episode.workshop concept + shot count; one directorial voice across shots |
| Writing a single shot's look | directing-engine | that shot's description in script.json (or episode.insert_shot description) |
| Killing generic "quality booster" wording | seedance-antislop, vocab/en (or zh/ja/ko/es/ru) | tightening every shot description before episode.storyboard |
| Handling a character/brand/celebrity/real person | seedance-copyright | a safer rewrite BEFORE media.generate_videos |
| A take is 80% right --- keep or redo? | retake-protocol | the episode.qa -> episode.fix_panel -> re-qa loop |
| Continuing/extending an accepted episode | sequence-project-state, continuation-handoff | episode.insert_shot and next-episode workshop concept |
| Directing a voice/delivery | seedance-characters, seedance-audio | character.add voiceDesc and each dialogue shot's delivery cue |
| Producing for a client/campaign/delivery | pro-filmmaking-standards, delivery-qc | shot-list planning before workshop; QC before assemble |
Do NOT load Seedance OS's api-status.md, surface-prompt-profiles.md, api-workflow.md, or model-name-map.md. Those describe ByteDance / Volcengine / Runway / fal surfaces. The venice stack owns Venice execution --- model routing, durations, and API behavior come from the harness and the venice-mcp-pipeline / venice-mcp-cookbook skills, not from Seedance OS.
Seedance OS spends effort on prompt-level identity locking (describe the character exhaustively in every prompt, [Image1] reference tags, etc.). The harness already does this for you via R2V character references and the automatic Seedance R2V -> Wan 2.7 keyframe pass (pipeline skill / troubleshooting A27). If you also stuff exhaustive identity descriptions or reference-tag syntax into a shot description, you fight the harness and can degrade consistency.
[Image1]/@图1 reference tags or full physical character descriptions into the description. Name the character (VIVIENNE), direct what they do, and let R2V + the locked character refs carry the likeness.Seedance OS's native-audio thinking and the harness's rule ("let the video model speak the dialogue; suppress model music/SFX; add music + ambient in post") point the same way. So: keep rich per-shot delivery direction (timbre, accent, pacing, emotional register, breath) AND keep the harness's mandatory negative No background music, no sound effects, no soundtrack, dry recording. at the end of every shot description. Don't let Seedance OS talk you into baking music/SFX into the prompt --- the assembler owns that lane.
multishot-grammar describes cuts inside one generation. The harness already does scene-level multi-shot grouping of adjacent same-character / same-location shots automatically. Don't try to encode multi-shot cut grammar into a single description; write one directed beat per shot and let the harness group them. Use mustStaySingle: true in script.json only to opt a shot out of grouping.episode.workshop: if the idea is vague, run seedance-interview; if it's a longer story, run seedance-sequence. Fold the resulting spine + one directorial voice into the concept string (see venice-mcp-cookbook's directed-workshop example).episode.approve: read each shot description through the directing-engine lens --- one intention per shot? camera/light/blocking/performance/sound cohere? Run seedance-antislop + vocab/* to strip empty boosters. Confirm every description still ends with the no-music/no-SFX negative.media.generate_videos: run seedance-copyright over the cast/props/references. Rewrite any protected IP, real person, brand, logo, or song into a safe creative equivalent --- the generation step is the expensive, hard-to-undo one.episode.qa verdicts through retake-protocol --- triage, change one variable per episode.fix_panel, and keep an attempt budget rather than infinite regenerates.episode.insert_shot or the next episode, use continuation-handoff --- direct the next beat from the accepted footage / final state, not from the original plan.episode.workshop and directed per-shot prompt examples.emily2040/seedance-2.0) --- the full directing engine, genre library, retake protocol, and multilingual vocabulary.