Let your agent write the prompt
A video model is not a chatbot: each one obeys a different grammar, and the same sentence that produces a clean oner on one model produces four cuts and an invented plot on another. These four agent skills write that grammar down — what each model actually obeys, measured on paid renders rather than guessed. Install them into Claude Code or Codex, describe the scene in plain words, and the agent hands you a prompt built the way that model wants it. They are the same skills this studio is run with, and they come with an account — which is free to open.
~/.claude/skills) and Codex (~/.agents/skills)Download the four skills (.zip)
Signing in is what the download asks for — an account is free, and the studio itself charges nothing to shoot on your own provider key.
What is in the pack
| Skill | What it writes |
|---|---|
seedance-clean | Seedance 2.0 / 2.0 Fast / 2.0 Mini — single shots and multishot sequences up to 15 seconds, in any visual style |
seedance-25 | Seedance 2.5 — up to 30 seconds, integer-second timing, up to 50 references, editing, extension, keyframes, storyboards, 3D blockouts |
gemini-omni | Gemini Omni Flash — one continuous shot instead of an auto cut-up, clip editing in conversation, extension to 40 seconds, readable on-screen text |
seedance-assets | Pre-production: four-view character sheets, location plates and prop sheets, with a locked style and an @tag registry the video prompts reuse |
Why a skill and not a prompt template
Because the useful part is the exceptions, and there are dozens of them per model. A few examples of what is written down in there:
- Seedance 2.0 does not understand timestamps. Writing “0–3s: he turns” wastes the line; 2.5 reads whole seconds and 2.0 does not. The skills route the brief to the right grammar instead of handing both models the same text.
- Gemini Omni cuts your scene up unless told not to. Left alone it invents several shots and a little plot; a single continuous take has to be asked for in so many words — the exact opposite of what Seedance needs.
- Style belongs distributed through the prompt, not stacked as a prefix of adjectives, and a character's emotion has to be written as observable performance, or the model averages it away.
- References are counted and given explicit roles. Every asset used gets one role; every asset not used gets listed as unused — otherwise the model quietly invents a use for it.
Install
The shortest route is to hand the job to the agent you are installing them for. Drag the downloaded archive into a Claude Code or Codex window — that gives the agent the file's path, wherever your browser put it — and ask: “Install the skills from this archive: unzip it into your skills folder (Claude Code: ~/.claude/skills, Codex: ~/.agents/skills), then tell me which skills you can see.” Both folders go in the sentence because the agent is the one that knows which of them is its own. A browser chat (ChatGPT on the web, claude.ai) can read an archive you attach but cannot install it — it has no access to your disk. The manual way is two lines:
-
Claude Code
Personal skills live in
~/.claude/skills:mkdir -p ~/.claude/skills && curl -sL https://longtake.studio/skills.zip -o /tmp/longtake-skills.zip \ && unzip -oq /tmp/longtake-skills.zip -d ~/.claude/skills
-
Codex
Same files, different folder — Codex reads
~/.agents/skills:mkdir -p ~/.agents/skills && curl -sL https://longtake.studio/skills.zip -o /tmp/longtake-skills.zip \ && unzip -oq /tmp/longtake-skills.zip -d ~/.agents/skills
To share them with a team, unzip into
.agents/skills/inside the repository instead — Codex scans that folder upwards from the working directory, and it survives a machine change. -
Ask for a scene
Start a new agent session first — skills are read when a session starts, so one that is already open will not see them. After that there is nothing to enable: the agent picks the right skill out of what you ask for. “Write me a Seedance 2.5 prompt — twenty seconds, a courier climbing a stairwell, dialogue at the door” is enough. To check they arrived, ask the agent to list the skills it can see.
Questions
- Do the skills work in Codex, or only in Claude Code?
- Both. They are plain SKILL.md files in the open Agent Skills standard, and Codex reads exactly that format — the only difference is the folder:
~/.claude/skillsfor Claude Code,~/.agents/skillsfor Codex, or.agents/skills/inside a repository for a whole team. Nothing in these four skills calls a Claude-specific tool, so behaviour is the same in either. - Do I need an account to download them?
- Yes — the archive is behind sign-in, because these skills are written from what this studio pays to find out. An account is free to open and free to shoot with: renders are billed by the model provider to your own key, at list price. Once downloaded the files are yours, they are plain Markdown, and nothing in them phones home.
- Is there anything for Veo or Kling?
- No, and deliberately. Veo takes ordinary prose — a skill would add ceremony without adding accuracy. Kling's character system needs assets registered on their side, which the studio does not do yet. When either changes, the pack changes.
- How do I get updates?
- Download the archive again and unzip over the top — the files replace themselves and nothing else in your skills folder is touched. The pack is rebuilt from the same skills used to shoot with here, so it moves when the models do.
What the prompts run on: