Seedance 2.5 Video Director turns video ideas into structured Dreamina prompts
Seedance 2.5 视频导演 Skill：提示词、脚本、诊断、人物锁定、续写、编辑、转场与首次使用指导
At a glance
- What is it?
- An MIT-licensed Codex skill that converts text ideas, reference images and old prompts into director plans, diagnostics and copy-paste prompts for Dreamina's Seedance 2.5. It only ever writes text: it never calls a paid generation API or spends credits.
- Who is it for?
- Seedance 2.5 Video Director suits Codex users who write Seedance 2.5 prompts often enough to want structure: locked characters, timed beats, and a conflict diagnostic before spending credits on Dreamina. It does not suit anyone expecting one-click video generation, because it deliberately never submits a task or touches credits.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 23 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 16, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
A planning layer for a video model that punishes vague prompts
Generative video models are unforgiving of sloppy instructions, and the README of this project says so plainly: prompts can only reduce identity drift, physics errors and camera conflicts, never guarantee identical results. Seedance 2.5 Video Director is a Codex skill for Dreamina, also known as 即梦, that works on that problem from the writing side. It takes text ideas, character reference images, videos, audio, storyboards or old prompts and turns them into executable video scripts, director plans, diagnostic results and prompts ready to paste.
Two boundaries define it. The repository is not affiliated with ByteDance, Dreamina or 即梦, and the skill produces text only: it never submits a generation task, never consumes credits and makes no promise about output.
What the skill covers, from 4-second clips to 180-second scripts
The feature list tracks Seedance 2.5's documented modes: ordinary videos of 4 to 30 seconds with an exact 30-second timeline, Long Video of 30 to 180 seconds, and continuation written before or after an existing clip. Editing modes are covered too, including Smart Edit, editing by marked region, and editing of an uploaded video. Render commands exist for the Clay Renderer's rough and fine white-model stages, transitions between two clips are a supported pattern, and multi-grid storyboard animation rounds out the list.
Character handling gets its own rule set: identity, body type, clothing and accessories can be locked; multi-character scenes, dialogue voices, BGM removal and audio-video sync are documented, alongside patterns for realistic emotional performance, couple scenes, competitive variety-show live shots and obstacle physics. The references/ directory adds prompt blueprints, official examples, capability notes and multimodal patterns, so the longer material sits in files rather than in one bloated SKILL.md.
Installing it into Codex
With Node.js present, one command installs the skill globally into the current user's Agent Skills directory:
npx skills add liyue-aigc/seedance-2-5-video-director -g -yThe -g flag means a global install and -y skips the confirmation prompt. The README notes that although the installer is launched through npx, the repository itself is not an npm package to add to project dependencies, so npm install is the wrong tool here.
Without Node.js, the documented path is to ask Codex to use its $skill-installer to install liyue-aigc/seedance-2-5-video-director from GitHub. Manual installation is a git clone into the skills folder, ${CODEX_HOME:-$HOME/.codex}/skills on macOS and Linux:
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills"
git clone https://github.com/liyue-aigc/seedance-2-5-video-director.git \
"${CODEX_HOME:-$HOME/.codex}/skills/seedance-2-5-video-director"Windows gets a PowerShell equivalent in the README, which defines the skills root under %USERPROFILE%\.codex\skills, creates it with New-Item, and clones into it. Afterwards you reopen the Codex session and call the skill by name in a prompt, in the pattern of the README's example: use $seedance-2-5-video-director to organise my video idea into a copy-paste Seedance 2.5 prompt. The installer shows a one-time learning guide on first use; later conversations go straight to the task, and you can ask for the tutorial again at any time.
Updates follow the same split: npx skills update seedance-2-5-video-director -g -y for installer users, a git pull inside the skills directory for manual clones.
First use: one message, one shot list
The README's quick-start example shows the shape of a request. It asks for a 20-second, 16:9 water-sports variety clip, names an attached image as the challenger's full-person reference with the original costume kept and the image background not inherited, lists the story beats from entrance to finish line, and specifies Mandarin commentary, crowd cheers, mechanism sounds and tense BGM with no subtitles. The requested output is a complete director plan plus the final prompt.
Two shorter modes exist. A prompt-only run asks for just the final prompt with no explanation, like the README's 15-second, 9:16 street-chase example with realistic handheld action style. A diagnosis run pastes an existing prompt and asks the skill to check only, without rewriting, whether time, camera movement, characters, physics and sound conflict with each other.
Character references follow a documented default: a full visible person, not just a face. Identity, facial features, skin tone and age impression are kept, as are hairstyle, height, body proportions and visible clothing. Background, pose, composition, crop, text and lighting from the reference image are not inherited unless you explicitly narrow the scope, for example by asking for a face-only reference.
A validator and twelve contract tests ship with it
The engineering hygiene is unusually visible for a prompt skill. Running python scripts/validate_skill.py locally checks the skill's frontmatter and name, required files and Markdown relative links, the first-use onboarding trigger rules, the invocation name in agents/openai.yaml, test case format with unique IDs and main pattern coverage, and leftover TODO or placeholder text. The same validator runs on every push and pull request through GitHub Actions.
tests/cases.json is a behavioural contract set: twelve cases in v1.0.0, according to that release's notes. It never calls a video generation service. Each case records an input scenario and expected output constraints, meant for manual regression or for an independent agent to test forward. That is an honest design for a text-only tool: the tests can assert structure, not generated pixels.
Limits the README states up front
The capability-boundaries section is worth reading before relying on the skill. It produces planning, scripts, diagnostics and prompt text, nothing else. Platform upload limits follow the current Dreamina or 即梦 interface, and the repository's capability summary is dated, so durations and modes can lag the platform. Because generation is random, better prompts reduce identity drift, physics errors and camera conflicts without ever eliminating them.
Maintenance looks healthy on the surface: v1.0.0 was published on 2026-08-18 and the last push was on 2026-08-26. Licence-wise the code and original skill content are MIT, while platform names, trademarks and third-party documentation remain with their owners, with sources explained in NOTICE.md.
Against prompting the model by hand
The alternative is writing Seedance prompts yourself or asking a general chat model to draft them. The difference is encoded defaults versus improvisation. This skill ships the full-person reference rule, the segmented timeline modes, Clay Renderer command patterns, a conflict diagnostic that checks time, camera, character, physics and sound, and a test suite that pins its behaviour. A general model improvises per conversation: it will not enforce the rule that reference images default to the whole visible person, and nothing checks its output for camera-versus-physics conflicts.
The trade goes the other way too. The skill's patterns are only as current as its notes on the platform's capabilities, while a hand-written prompt can react to a UI change the same day. For repeated, structured work the skill wins; for a one-off experiment right after a platform change, direct prompting may be quicker.
Editorial conclusion
Seedance 2.5 Video Director suits Codex users who write Seedance 2.5 prompts often enough to want structure: locked characters, timed beats, and a conflict diagnostic before spending credits on Dreamina. It does not suit anyone expecting one-click video generation, because it deliberately never submits a task or touches credits. Before installing, check that the capability notes in references/capabilities-and-limits.md still match what the Dreamina interface offers, since the README warns the platform moves. The last push was on 2026-08-26, the validator and contract tests run without network access, and the MIT licence covers the skill text but not the platform names it orchestrates.
Frequently asked questions
How long can Seedance 2.5 videos be?
Per the skill's capability notes, ordinary videos run 4 to 30 seconds with an exact 30-second timeline mode, and Long Video runs 30 to 180 seconds. The skill writes scripts for both ranges.
Does seedance-2-5-video-director generate videos or spend credits?
No. The README states it only produces plans, scripts, diagnostics and prompt text, and it never submits generation tasks or consumes credits.
How do I update the skill after installing it?
If you installed through npx skills, run its update command for this skill; manual clones update with a git pull inside the skills directory. The README documents both paths for Windows, macOS and Linux.
Community notes