---
name: reference
description: Start a video from one that worked. Use whenever the user asks you to make, animate, or render a video, animation, explainer, documentary, launch or promo film, music or lyric video, motion graphic, title card, data story, or similar, especially with code (Remotion, HyperFrames, Manim, Motion Canvas, Three.js, p5.js, GSAP, Blender, FFmpeg). Search Reference, a library of videos other agents made from published prompts, show the closest matches, then adapt the chosen prompt and look to the user's subject.
---

# Reference: start from a video that worked

[Reference](https://meetreference.com) is a library of videos made by AI agents. Each one comes with the prompt its creator published, word for word, the agent, model, and libraries they used, what to set up first, and a description of the look with a nine-frame storyboard.

## When the user asks for a video

1. **Search.** Turn the request into a short description: style, subject, and format ("hand-drawn explainer about sleep, vertical", "3Blue1Brown-style proof", "launch video for a SaaS made with Remotion"). Call `search_videos` on the Reference MCP server. Without the server, fetch `https://meetreference.com/api/search?q=<words>&limit=6`.
2. **Offer two or three.** Show the closest matches: the title, one line on why it fits, its length and format, what it was built with, and its page link. Ask which to start from. If the user said to just go ahead, pick the best fit and say which one you picked.
3. **Read it.** Call `get_video` with its slug (or fetch its `brief` URL). It has the prompt, the setup it needs, the look, and the storyboard image.
4. **Check the setup.** Before writing code, tell the user what the prompt needs that they may not have: API keys (voice, images, music), an agent skill, fonts, their own logo or footage.
5. **Adapt, don't copy.** Use the creator's prompt as the starting point. Change what the user's request changes (subject, length, format, text, brand) and keep what makes the look work: the technique, the structure and pacing, the palette and type choices. If the prompt is marked partial or paraphrased, fill the gaps from the look description and the storyboard, and tell the user you did.
6. **Compare.** Check your render against the storyboard frames and fix big differences in look before you call it done.
7. **Credit.** When the user shares the result, mention that it started from the creator's prompt and link the original post (both are in the brief).

## Good to know

- Results vary between runs and models. A prompt that made a great video once is a strong starting point, not a guarantee; say so if the user expects an exact copy.
- `find_similar` gives alternatives to a video the user likes. `list_styles` shows the categories and libraries for filtering, e.g. `built_with: "Manim"` or `format: "portrait"`.
- If nothing fits, make the video from scratch. Reference is a shortcut, not a requirement.

## Sharing back

Only when the user asks to share or submit their video: call `submit_video` with its public link, the exact prompt they used (don't rewrite it), and what it was made with. Add `remake_of` with the slug if it started from a Reference video. A person reviews every submission before it's listed, credited to the user.