How To Create 3D Documentary Videos With AI (Step By Step)

A documentary channel called Loom pulled over 6 million views in 30 days (estimated around $17K that month) and more than 48.5 million in total. Black Files did 1M+ views in a month. They all use the same 3D mannequin animation style to tell stories. Most tutorials throw a pile of complicated tools at you. This one keeps it beginner-simple and shows the trick that keeps your character identical in every scene.

Step 1: Prime your chatbot

Any chatbot works (Claude, DeepSeek, Gemini, ChatGPT). What matters is how you talk to it. Open my free Google Doc, copy Chapter 1, paste it into Claude and add “Here’s the task.” This tells the AI exactly which niche you’re in.

Step 2: Give it the niche PDF

Claude asks for a source PDF with channel screenshots, branding references, transcripts, thumbnails and image templates from channels winning right now. Get it from Chapter 2 of the doc, attach it, and wait for “PDF absorbed.”

Step 3: Ideas and script

Type generate video ideas to get 10 ideas built on proven patterns. I used The evil design of America’s private prisons. Pick a length (1–10 min; 20–30 min is the sweet spot for this format), then ask for a tone rewrite if you want it more engaging or casual.

Step 4: Voiceover

In ElevenLabs, use v3, paste the script, and click Enhance before generating. That one step makes it sound far more natural.

Step 5: The consistency trick, part 1: characters

Type next and Claude writes character prompts based on your script. In Higgsfield (50% off) → Image → Nano Banana Pro, generate each character.

Step 6: Story chapters

Type next again. Describe your own visual flow (best) or type design it. Claude split my script into five chapters of about 12 seconds each.

Step 7: The consistency trick, part 2: Seedance 2.5 with references

I tested the free video tools and none matched the quality, so I used Seedance 2.5 on Higgsfield (on desktop, not phone):

• Attach your character images in the reference section. Don’t skip this.
• Paste the full chapter prompt and set the duration to match (e.g. 12 s).
• Seedance 2.5 does up to 30 s in one shot, takes up to 50 references, and goes up to 1080p. Unlimited mode at 720p is great for testing.

Repeat for every chapter, keeping the references attached.

Step 8: Edit

Voiceover first, cut dead air, then match clips to the narration. No clip for the opening line? In CapCut use Text effects, search date, pick a template, make it a compound clip, and add Glow and an old TV effect. Speed clips up or down to sync, and add suspenseful music from the YouTube Audio Library.

Step 9: Thumbnail

Ask Claude: “I want a thumbnail for this video,” generate the prompts in Higgsfield, and pick the best.

📝 Your notes on this video

Saved on this device only

Copy notesDownload .txt