Make 3D AI Animated Videos Like Ifverse With Claude Code

Prompt 1: install everything (copy all of it)

Set up this folder as a workspace for making animated 3D faceless videos with code. Install everything below, check each tool works, and render a short test video at the end.

Do every step you can yourself. Check what is already installed before installing anything, and don't reinstall what works. If you need me to do something (an admin password, a sign-in, a download only I can do), stop and give me short, simple, step-by-step instructions, then wait until I say I'm done.

**1. Base tools**
- Node.js 22 or newer, Python 3.11 or newer, git, and FFmpeg (with ffprobe).
- yt-dlp, to download the reference video for analysis.

**2. Video engine and 3D**
- HyperFrames (HTML rendered to video). Create the project with `npx hyperframes init`, then run `npx hyperframes check` until it passes.
- Install the HyperFrames skills for Claude Code with `npx hyperframes skills update`.
- Three.js for the 3D scenes, with its GLTF loader so it can load 3D models.
- GSAP for the animation timelines.

**3. Analysis tools (for the reference breakdown)**
- Shot and cut detection: OpenCV and PySceneDetect.
- Word-level transcription: WhisperX (or faster-whisper with word timestamps).
- On-screen text reading (OCR): EasyOCR or Tesseract.
- Audio measurement: librosa and pyloudnorm.
- Contact sheets and frame grids: Pillow.
- Put the Python tools in a virtual environment inside this project, not system-wide.

**4. Assets**
- Make folders for `assets/models`, `assets/characters`, `assets/sfx`, `assets/music` and `assets/voice`.
- Use only free assets I can use commercially. Prefer CC0 3D models (for example Kenney or Poly Pizza) and CC0 sound effects. Keep a `CREDITS.md` with every asset's source, author and license.
- Rigged, animated characters: Mixamo needs a free Adobe sign-in, so give me the steps to download what you need and tell me where to put the files.

**4b. Claude Code add-ons**
- vidIQ for YouTube research: run `claude mcp add --transport http vidiq https://mcp.vidiq.com/mcp`, then tell me to run `/mcp` and sign in with my vidIQ account (the free plan works).
- The Humanizer skill (github.com/blader/humanizer) for rewriting scripts so they don't sound like AI: `claude plugin marketplace add blader/humanizer` then `claude plugin install humanizer@humanizer`.
- Sound effects: clone github.com/Bodila51/opus-sound-layer into this folder, install `uv`, and build its sound library with `uv run scripts/sfx.py kit` (runs in the background, about 300 MB).

**5. Voice**
- I record the narration myself or use a text-to-speech voice. Set up `assets/voice/` and a script that turns the voice file into word timestamps (`words.json`) with WhisperX.

**6. Test it**
- Build a 5-second 3D test scene in HyperFrames: a small Three.js world with a model, a light, a slow camera move, and one object that animates in on a word from a short test voice clip.
- Render it to MP4 at 1920×1080, 30 fps, and show me a frame from the start, middle and end.

**7. Report**
- Give me a short table: each tool, its installed version, and pass/fail.
- Write a `SETUP.md` in the folder with the same list and the commands to start a new video, so I can redo this on another computer.

Prompt 2: break down a reference video (copy all of it)

Watch and analyze this reference video frame by frame: LINK HERE

Run the full frame-by-frame breakdown. Watch it with sound, build a time-coded shot list, and inspect the frames immediately before, during and after every visual change. Do every step you can yourself. If you need me to do something, give me short, simple, step-by-step instructions and wait until I finish.

**1. Get the source**
- Download the video and save its info (title, channel, upload date, views, length, resolution, fps).
- Extract the audio, every frame at half size, the first 10 seconds at full size, and a 1-frame-per-second contact sheet.

**2. Map the whole video**
- Map the full length at 2 fps and label every shot by type (wide set, close-up, character, prop insert, diagram, data card, text card, transition).
- Run an automatic shot/layout classifier, then check its output against the frames by eye and fix what it got wrong.

**3. Cuts and camera**
- Find every instant visual change. Check a before/after pair for each one by eye.
- Separate a hard cut to a new shot from a camera move inside the same shot (push-in, pull-out, pan, orbit, dolly).
- Measure every camera move: start and end time, duration, scale change (for example 0.8 s, ×1.095), direction, and easing.
- Record shot lengths: median, shortest, longest, and how they change between hook, body and ending.

**4. Motion in seconds 0–10**
- Inspect every object, character, icon and text movement in seconds 0–10: where it enters from (left, right, top, bottom, depth), how long it takes, how far it moves, how much it scales, how long it stays, how it leaves, and its easing.
- Note 3D details: camera lens and angle, lighting and shadows, materials, depth of field, how characters are animated (walk, idle, gesture), and what moves in the background while nothing else happens.

**5. Sync to the voice**
- Transcribe the narration with word-level timestamps.
- For every cut and every new object, find the first spoken word of that beat and check that the visual starts on that word's audible onset. Report the offset in milliseconds.
- Mark where beats start: at sentence starts, or mid-sentence on the noun the next shot shows.

**6. On-screen text**
- Read all on-screen text (OCR): labels, numbers, titles, captions. Record when each appears and disappears, its position, size, style and animation.

**7. Motion density (graphics never land and sit)**
- Measure how much of the time something on screen is moving, how many new visual events start per second, the median gap between events, and the longest stretch where nothing new happens.
- List every moment where the picture lands and waits while the narrator keeps talking, and what the video does instead to keep it alive (slow camera drift, ambient loops, background life, the next action starting before the last one settles).
- Describe how one idea grows out of the last one (morphs, camera flying to the next place, objects transforming) instead of the screen emptying and rebuilding.

**8. Sound**
- Measure loudness, music level under the voice, and every sound effect hit: its time, its type (whoosh, pop, impact, riser, ambience), and which visual it belongs to.

**9. Write it up**
- **ANALYSIS.md**: the evidence. Source facts, how each thing was measured, the shot list, camera measurements, the 0–10 s motion inventory, sync offsets, text, density numbers and sound, with links to the contact sheets and data files.
- **METHODOLOGY.md**: a short, deterministic editing method I can reuse: clear if/then rules for beats, cuts, shot types, camera patterns, motion presets, word-anchored timing, on-screen text and sound; density targets taken from the measured numbers (the new video must match or beat them, with no stretch where the graphics land and sit); and a final QA checklist to run before every render.
- If ANALYSIS.md and METHODOLOGY.md already exist, add the new patterns as new sections and keep what is already there.

**10. Then replicate it**
- Confirm the setup from the tools prompt still works: render the test scene again.
- Wait for my OK on the breakdown. Then rebuild this style for my video with my own story and information, following METHODOLOGY.md. Use your own characters, sets and assets, not the reference's.

Prompt 3: video ideas with vidIQ (copy all of it)

Use vidIQ to research faceless animated "What If" story channels on YouTube, such as Ifverse, Monk Mode Films and Day One Films, plus any similar channels vidIQ finds.

1. For each channel, pull subscribers, total views, how fast it grew in the last 30 days, and its 10 most-viewed videos with their titles and view counts.
2. Find the patterns in the titles that get the most views: the shape of the sentence, the length, numbers, the type of "what if", and which topics repeat.
3. Give me 20 new video ideas for my channel. Each idea is a title plus one line on the story. They must be my own stories, not copies or remixes of theirs.
4. Score each title with vidIQ and sort the list from best to worst.

Stop and let me pick one before you write anything else.

Prompt 4: the script (copy all of it)

Write a 1-minute narration script for the title I picked: [TITLE].

- About 130–150 words. It must read as one voice, a narrator who speaks to "you".
- Open on a concrete moment the viewer can see, not a question or a big statement.
- Short sentences. Plain, everyday words. Use dialogue between characters and give each line its own sentence.
- One clear turn in the middle, where the power or the situation starts to cost something.
- End on a quiet, personal moment, then one final question to the viewer.
- Every sentence must be something I can show on screen. Name places and objects.
- Don't use: "Imagine", "In a world where", "little did you know", "but here's the thing", "it's not just X, it's Y", "let that sink in", or any closing line that explains the moral.
- Give me plain text only, with no labels, speaker tags or stage directions, so I can paste it straight into a voice tool.

Prompt 5: humanize the script (copy all of it)

Run the script through the humanizer skill. Fix every sign of AI writing it finds, then show me the final script and a short list of what changed.

**To make the humanizer better over time:**
Read Wikipedia's "Signs of AI writing" page and other recent lists of AI writing tells. Compare them with the patterns already in the humanizer skill. Show me the tells it doesn't cover yet, with one example each, and add the ones I approve to the skill.

Prompt 6: build the video from your voiceover (copy all of it)

Here is my voiceover: [voiceover.wav]. Build the video in the style we set up from the reference.

1. Transcribe the voiceover into word timings.
2. Make a shot list from the words. Each new line or idea gets a new shot, the cut lands just before its first word, and the camera moves in every shot.
3. Build new worlds from this script. Before you build, list every place and object each line names. Create those places, and new worlds where the script needs them, so that every visual matches what is being said. Don't reuse the reference's world or any earlier video's world unless the script asks for it.
4. Fill every world with high-quality detail so it never looks empty or boring: furniture and props, set dressing, textures, background life, and lighting that sets the mood.
5. Add visual effects: light beams, floating dust, glows, particles, and an effect for the story's main idea. Characters react on the words that carry emotion with blinks, expressions and gestures.
6. Add chapter labels and the end-card question the way the reference does.
7. Check every shot by eye, render a draft, run the motion check, and show me a contact sheet of every shot before you call it done.

Prompt 7: fix and improve (copy all of it)

Look at this video like a master 3D motion designer. Go through it shot by shot and make it better:

- Make sure every visual matches what the voice says at that moment.
- Fix anything empty, dark, blocked, cut off or hard to read.
- Add more detail and life to the worlds, and better visual effects where they help the story.
- If it helps, change the color theme or the look of a world.
- Here are my notes: [YOUR NOTES, e.g. "the bedroom is too dark", "make the rain only in the rain shot", "make the city busier"].

Show me a before and after contact sheet of the shots you changed, then render again.

Prompt 8: sound effects (copy all of it)

Add sound effects with the opus-sound-layer repo in this folder. Read its PROMPT.md and follow it.

Place the sounds from the video's own timing: cuts, character moves, things appearing, and the big moments. Pick every sound yourself from the kit, and keep one sound family for the whole video. Mix them with my voiceover, run its QA check, and give me the final MP4 plus 5 timecodes I should listen to.

Prompt 9: YouTube assets (copy all of it)

The video is done. Write the YouTube assets for it:

1. 10 titles: 5 that follow the "What If…?" shape of the top channels and 5 alternatives. Score them with vidIQ and mark your top 3.
2. A short description of 2–3 sentences that makes people curious without spoiling the story or the ending.
3. Tags: 15–20, ordered from most to least relevant.
4. Thumbnail text: 5 options of 1–4 words each, different from the title.
5. 5 thumbnail ideas based on the title. For each one: the main image, the character's expression, the colors, and where the text goes.

Prompt 10: turn it into a skill (copy all of it)

We now have a video I'm happy with. Turn this whole workflow into a Claude Code skill, so next time I only give you a video idea and you do the rest the same way.

1. Use the skill-creator skill from Anthropic (github.com/anthropics/skills). Name the new skill `3d-story-video` and save it in `~/.claude/skills/`.
2. Put everything we built and learned into it:
   - the steps in order: ideas with vidIQ, script, humanize, wait for my voiceover, build, fix, sound effects, YouTube assets;
   - the methodology and style files from the reference breakdown;
   - the 3D kit: world code, character rig, props, effects and the render settings;
   - every fix I asked for along the way, written as rules so I never have to ask again (for example, "build every world from the script" and "never reuse an old world").
3. Keep my approval stops: after the ideas, after the script, after the first render, and before the final export.
4. Test the skill on a new idea: run it up to the script, then build a short test from a voiceover I give you. Show me the result and anything you changed in the skill.