AI Video Prompts

Real prompts behind viral AI-made videos — collected across Claude, Muse AI, ChatGPT, and Gemini. Every entry links the creator's original post so you can see the result and credit the source.

Two genres live here side by side:

• Code-driven animation — the prompt asks the model to write the animation as code (HTML, Canvas, SVG, Three.js, GSAP, shaders). Dominates the Claude/ChatGPT/Muse AI entries. Render with any browser, or pipe through HyperFrames for video output.
• Video-gen prompts — the prompt is the video direction (Veo, Sora, and similar). Common in the Gemini entries. Paste into the matching video model.

What's inside

model	entries	mostly
claude	475	code-driven (Opus 5.5)
muse-ai	7	official video-gen prompts + early creator clips
chatgpt	24	code-driven + Sora 2
gemini	39	Veo 3 video-gen

Categories (shared across models): explainer (diagram builds, concept visualizations, data-viz), motion (kinetic type, particles, showreel-style motion graphics), 3d (Three.js/WebGL/GLSL scenes), interactive (playable pieces, physics toys).

Top tech tags: canvas, svg, threejs, shader, gsap, css, ai-video, veo, sora.

Layout

• data/videos.json — machine-readable index (545 entries). Fields: slug, model, model_version, title, author, author_url, post_url, category, tech_tags[], prompt, prompt_partial, added, featured.
• data/provenance.json — per-entry curation notes (why it's notable, which source table it came from) for the 70 hand-curated entries.
• prompts/<slug>.md — the prompt text as the creator shared it, one file per entry.

How to use

1. Find candidates. Filter data/videos.json by model, category, tech_tags:

   python3 -c "
   import json
   vs = json.load(open('data/videos.json'))
   for v in vs:
       if v['model']=='gemini' and 'veo' in v['tech_tags']:
           print(v['slug'], '-', v['title'])
   "
   

2. Read the prompt. Open prompts/<slug>.md (or the prompt field). Note prompt_partial: true — the creator only shared part of it. Treat those as inspiration, not recipes.
3. Adapt, don't paste blindly. Swap the subject matter, keep the technique: how the prompt describes motion, timing, camera, layers, loop structure. Technique transfers across models better than literal wording — a Veo prompt's shot language ("slow dolly-in, 35mm") works in Sora too; a canvas prompt's easing language works in any code model.
4. Credit the creator if you publish something visibly derived — post_url is on every entry.

Model notes

• Claude (Opus 5.5): the largest set, all code-driven. Prompts lean cinematic and precise — good technique mine for any code model.
• Muse AI: small seed set — 4 official Meta preview prompts plus early creator clips. Expect this section to grow via community submissions once Muse Video goes public.
• ChatGPT: two flavors — code-driven builds (GPT-6 Astra writing games, 3D scenes, cartoon pipelines) and Sora 2 video-gen prompts (post text is the prompt).
• Gemini: Veo 3 video-gen prompts from the launch wave — shot lists, camera moves, subject choreography. Model-specific in knobs (Veo camera syntax) but the shot design transfers.
• Some prompts reference paid tools or MCPs — skip or substitute those parts; the core technique is the value.

Attribution

• Claude entries (475) adapted from yihui-dev/awesome-opus5-5-videos (MIT).
• Muse AI entries (7): 4 official Meta preview prompts via imaginevid/awesome-muse-video-prompts-and-skills (CC BY 4.0); 3 curated originally.
• ChatGPT entries (24) selected, re-titled, and re-tagged from kikichalks/awesome_sora2_prompt (MIT) and beatapi/awesome-3d-prompts (MIT).
• Gemini entries (39) selected and re-tagged from akirakai/awesome-veo3-videos (CC0).

Every entry credits its creator via post_url — the prompts are their words, quoted with credit. "Viral" is survivorship bias: these are the hits. Use them for craft, not as a guarantee.