venice-video-harness: Skill for Claude Code

Seedance 2.5 Prompt Optimizer

Purpose

Compile the user's raw text, novel excerpts, and optional image, video, and audio references into one clean Prompt that can be submitted directly to Seedance 2.5. Preserve the user's core intent while making material roles, subject mappings, event states, and relationships to preserve explicit.

This Skill's responsibility ends with understanding the input and producing the Prompt. It must never invoke a generation API by itself. When the user explicitly asks to generate a video, first use this Skill to produce a clean Prompt, then pass it to a separate generation tool or workflow. This Skill always treats reference materials as read-only: it does not modify source materials or automatically create auxiliary materials.

When to Use

Load this Skill when:

• The user asks to optimize, rewrite, or complete a Seedance 2.5 Prompt.
• The user provides a rough idea, long-form plot, novel excerpt, or unstructured complex Prompt.
• The user provides images, videos, audio, file paths, or another multimodal request that requires material-role mapping.
• The user wants multi-reference generation, a long video, keyframe control, a storyboard grid, blockout rendering, video editing, audio editing, or video extension.
• The user wants to improve emotional performance, camera movement, audio, dialogue language, or a product/process demonstration.

Do not load this Skill automatically when:

• The user only asks about API parameters, pricing, quota, errors, or model capabilities and does not need a Prompt.
• The user only asks to evaluate a generated video and does not ask for a rewritten Prompt.

When the user explicitly asks to generate a video, still compile the raw input into a clean Prompt with this Skill before passing it to the downstream generation workflow. This Skill does not submit the generation task itself.

Non-Negotiable Principles

1. Intent first: Do not change character identities or counts, key props, scenes, event causality, spatial relationships, the edit target, extension direction, or story outcome.
2. Template first: Reorganize every request with the template for the current task, even when the original Prompt is already complete. Do not merely paraphrase it.
3. One explicit role per material: State what to use from every material that is actually activated. List every available but unassigned material individually under [Unused Materials] so downstream PE does not reactivate it.
4. User mappings first: Never override material roles, subject names, or relationships explicitly assigned by the user. Once a slot is covered, do not add an unmentioned material merely to reinforce the same role.
5. Ask as little as possible: Resolve anything that can be reasonably inferred from the text, materials, and context. Ask one consolidated question only when ambiguity could change the core result and several equally plausible interpretations remain.
6. Submit-ready content only: The Prompt must not contain analysis, evaluation notes, experiment groups, model versions, run labels, API keys, or reasons for the rewrite.
7. Separate parameters from creative content: Do not write aspect ratio, total duration, resolution, frame rate, or the audio toggle into the Prompt. For ordinary generation, set them on the generation page or through the API. For video editing, first-frame or first-and-last-frame generation, and video extension, follow the task's automatic parameter-locking rules. Time ranges already written by the user are creative content and must be preserved; do not invent new numeric time ranges merely from a target duration supplied by the page or API.
8. Do not add blanket constraints: Do not automatically add unrequested quality or stability boilerplate, watermarks, logos, subtitles, duplicate-subject restrictions, or other generic negative constraints.
9. Return one best version: By default, output only the single Prompt judged most appropriate. Return multiple versions only when the user explicitly asks for a comparison.
10. Separate facts from observations: Character identity, age, relationships, events, and outcomes come from the user's text. Materials may contribute only directly visible or audible attributes; never promote a visual guess into a story fact.
11. Match subject cardinality: A single-person character sheet must not define two named characters who appear at the same time. When several single-person candidates are available, first assign one image per person; when several group candidates are available, first assign one image per group. Combine references only for multiple views of the same subject or when the material itself clearly contains the same story group.