StewdioAI/blank-page

What it does

blank-page is a skill for Claude Code, Codex and other coding agents that turns a vague motion-graphics idea into a detailed generation prompt. The problem it targets is the empty prompt: a one-liner gives the agent too little to aim at, so it defaults to a safe, generic result. The skill instead searches for visual and motion references, writes a numbered beat sequence, and attaches real assets before returning the expanded prompt README lines 3–20.

It is aimed at people working on short motion pieces — launch clips, social posts, website backgrounds — who know the purpose but not the look, or who want directions they would not have thought of. The author is Blake Shao, and the project is connected to Stewdio, a prompt-based motion-graphics editor Stewdio.

Quick start

Install the skill with the skills CLI from the repository root or any working directory:

npx skills add StewdioAI/blank-page

The installer asks which coding agents to install for; add -g to install for every project rather than only the current one Install.

Then invoke it in Claude Code with /blank-page (Codex uses $blank-page) and describe the clip conversationally:

/blank-page a 5 second launch video for my app

You can also add purpose, mood or a brand reference, or say "expand this prompt" / "help me write the prompt for ..." to trigger the same behavior How to use. The result is a full prompt with references attached and up to three creative calls you can swap before passing it to an agent.

The README's before/after pairs use the same one-line prompt run directly versus through /blank-page, with generated clips shown side by side before/after examples. More worked examples are in references/examples.md.

How it works

The skill follows a four-step procedure How it works:

1. Pick a look — five images from one aesthetic on are.na, Pinterest or the open web; a named style or brand keeps the search inside it.
2. Find the motion — three gifs or videos whose movement can be referenced.
3. Write the beats — a numbered sequence built from those references, with at least one twist such as a reveal, morph or zoom-out.
4. Add real assets — a 3D model, footage or the real UI instead of a hand-built stand-in.

Short edit notes such as "make it faster" skip the search and return a precise note instead. The repository layout puts the procedure in SKILL.md, search/download/sheet scripts under scripts/, and reference material under references/ Layout.

Requirements and constraints

• python3 (standard library only) and curl are required for the reference scripts.
• ffmpeg is needed only for contact sheets (find_refs.py sheet, channel <slug> <dir>).
• Pinterest pin discovery relies on web search being available in the agent.
• SKETCHFAB_TOKEN is optional and only needed for Sketchfab downloads not covered by the Objaverse mirror Requirements.
• The skill is designed to work with any coding agent, but the README specifically documents invocation for Claude Code and Codex.
• Reference searching pulls media from external sites, so availability and licensing of those images, gifs and models are governed by their sources, not by this project. The repository itself is MIT licensed LICENSE.

HotRepo's take

The repository is new and small, and the recorded data is limited. It appeared on HotRepo's new Python list on 2026-11-05 with 108 stars, gaining 7 stars in a roughly one-hour window, and had 2 forks; no all-language daily rank has been recorded repository. That single snapshot is enough to say the project drew some early attention, but not enough to describe a trend or adoption level. The more substantive signal is the design: it is a concrete, inspectable agent skill (procedure, scripts and references) rather than a wrapper around a hosted service, and the README shows before/after outputs produced by the same model with and without the skill.