Video Prompting Skill
Draft and refine format-compliant prompts for ten-plus major video generation models, plus a character-consistency workflow.
This is a pure text prompt-drafting guide with no scripts, network calls, or sensitive-data handling; least privilege is trivially satisfied. README provides repository attribution and Apache-2.0 license, and each model guide cites official sources, giving transparent data flow. Deductions: publisher unverified, no rollback/recovery mechanisms described.
SKILL.md's routing logic (model index, workflow index, formatting rules) is internally consistent; reference paths match the index, and the MiniMax H3 mode/schema/timestamp rules are exhaustive with a final compliance check serving as controlled failure feedback. Deductions: static review cannot execute anything, no test suite or CI evidence is shown; several referenced files (character-sheets.md, Seedance, Wan, Veo guides) are absent from the material, so their completeness and error handling are unverified.
The description lists explicit triggers (model names, 'video prompt', 'character sheet prompt') and modes (t2v/i2v/reference-driven); scenarios and boundaries are clear, including fine distinctions like complete silence vs. no-music-only. Deductions: core value depends on overseas services (video-model platforms, docs at X/HuggingFace/ltx.io), potentially unreachable from mainland China; no Chinese-language support declared.
Good progressive disclosure: SKILL.md entry point routes to references/ subdocs, with documented extension patterns; README covers install, .skill bundling, and a tutorial. Deductions: no version number/changelog, no FAQ or top-level known-limitations section (weaknesses only embedded per model), maintenance ownership and update path unstated, publisher identity unknown; README is truncated mid-file.
The goal (model-schema-compliant video prompts) is clearly structured with weak/strong example rewrites and full example files (MiniMax H3 full-reference, motion-graphics trailer, LTX-2 example set); field-level MiniMax H3 schema rules offer real marginal value over manual drafting. Deductions: no execution evidence that outputs are directly usable; effectiveness is inferred from in-document examples only.
Each model guide cites official sources (LTX blog/docs, MiniMax HuggingFace guides, GitHub repos), with fact/inference separable and example files available for independent comparison. Deductions: source links not verified reachable or consistent during static review, coverage limited to shown files, and no third-party execution corroboration.
- Publisher is not verified by the FollowSkills registry; identity is unknown — verify repository provenance before use.
- This is a static source review; nothing was executed and output quality has not been validated against actual video-model generation.
- The skill depends on overseas official docs and video-model platforms that may be unreachable from mainland-China networks.
- Several indexed reference files (character-sheets.md, Seedance, Wan, Veo guides) were not shown in the provided material; their completeness is unverified.
- No version number or changelog, making updates and changes hard to track.
What does this skill do, and when should you use it?
This is a pure-prompt Agent Skill that turns a user's creative intent into ready-to-use video generation prompts, or into image-model character-sheet prompts for consistency before image-to-video. It ships model-specific prompting guides for MiniMax H3, Seedance 2.0/2.5, LTX-2/2.3/2.5, Veo 3/3.1, Wan 2.2, Wan Animate 2, and Ovi, routing each request to the right guide and following its format rules strictly. A character-sheet workflow (turnarounds, expression sheets, photoreal identity sheets) supports a recommended handoff order: character sheet → scene still → video prompt. There are no scripts or runtime dependencies, making it a low-friction choice for creators producing video prompts across models.
Reads SKILL.md plus the relevant references/models/*/prompting.md guide; first routes the request to either a video prompt or a character-sheet prompt; confirms model and input mode (t2v, i2v, reference-driven, first/last-frame, etc.); drafts the prompt following the selected guide's section order, dialogue/audio format, and shot structure; on output, strips model name, duration, resolution, and aspect ratio from the prompt text, supplying parameters separately in a recommended-parameters line; for i2v, treats the image as the visual anchor and focuses text on motion, camera, emotion, and audio.
What are this skill's strengths and limitations?
- Covers model-specific prompting rules for ten-plus mainstream video models rather than one-size-fits-all
- Full character-consistency pipeline: character sheet → scene still → video prompt
- Pure Markdown instructions with no script dependencies — trivial to install and audit
- Careful output rules: single-line default, parameters separated from prompt text
- No automated tests or benchmark data; guide quality must be verified yourself
- Tied to specific model versions (e.g., Wan 2.2, LTX-2.5) that may age as models update
- No generation examples or user feedback provided in the source
- Effectiveness depends on the host agent's ability to follow long instructions; weaker models may comply partially
How do you install this skill?
Easiest: give the repo link https://github.com/Square-Zero-Labs/video-prompting-skill to your AI agent and ask it to use the skill. Claude: import via folder — select video-prompting-skill/video-prompting/ — or build a .skill zip bundle per the README. Codex: user-scoped install with mkdir -p ~/.codex/skills && cp -R video-prompting-skill/video-prompting ~/.codex/skills/video-prompting, then restart Codex; or place it in .codex/skills/ for repo scope.
How do you use this skill?
Prompt your agent with something like: 'Use the video prompting skill to create a prompt for MiniMax H3 in which ...' — or trigger via model names (Seedance 2.5, Veo 3, etc.) or phrases like 'character sheet prompt' / 'photographic identity sheet'. The skill routes the request, asks minimal clarifying questions, and outputs the final prompt; output is single-line by default, with 2–3 variants on request.