Digital Human Koubo: AI presenter videos from one photo
Its author turned one photo and one sentence into a 27-second livestream ad for 482 credits.
- Who
- Published under MIT by Berryxia (andyhuo520) as an agent skill. The README says the repository itself was built with Claude Code.
- When
- The repository opened on October 26, 2026, and v1.0.0 with the demo videos came out the same day.
- Where
- Tested in Claude Code; Codex, OpenClaw and WorkBuddy should run the same scripts. Generation on Renoise's cloud, editing on your machine.
What was hard
AI video models make about 15 seconds at a time. A video of several minutes meant writing a prompt per clip, listening to each one to check it said the script, evening out voices that each came from a different generation, and then laying captions and graphics on top.
What it does about it
The skill turns that into steps an agent follows. It cuts the script into segments, quotes the credits first, generates only the first segment once you agree, and batches the rest after you have checked it. Every clip is compared with the script by whisper, the voices are matched so thirty clips sound like one recording, and the chosen style and 48 packaging components are rendered in Remotion.



How it runs inside
- 1
You → Claude Code
paste content → 2 styles suggested → brief.jsonA one-line diagnosis suggests two of five styles with previews, then asks ratio, length, captions and presenter at once.skills/digital-human-koubo/SKILL.md - 2
Claude Code → build_tasks.py
script.json (≤ 15 s per segment)Each segment's length is characters ÷ 5.4 per second + 1 s; any segment over 15 s stops the run until it is split.skills/digital-human-koubo/scripts/build_tasks.py - 3
build_tasks.py → You
segments · seconds · estimated credits720p costs 17 credits a second, 480p picture-in-picture 10; nothing is generated until you agree.skills/digital-human-koubo/scripts/build_tasks.py - 4
Claude Code → Renoise MCP · Grok Video 1.5
tasks.json in batches of 10 → downloadTasks carry the reference photo and prompt; rangedl.py fetches each clip in 6 resumable ranges, then ffmpeg decodes it.skills/digital-human-koubo/scripts/rangedl.py - 5
Renoise MCP · Grok Video 1.5 → assemble.py · whisper
renoise-<taskId>.mp4whisper transcribes each clip and flags it under 85% match; it keeps 0.25 s after speech, joins clips at -16 LUFS.skills/digital-human-koubo/scripts/assemble.py - 6
assemble.py · whisper → MP4 rendered in Remotion
tl.json word timeline → renderalign.py times every character, and the style template and pack components appear on those times.skills/digital-human-koubo/scripts/align.py
What goes in, what comes out
Is it for you
Good for
- Posting short or explainer videos regularly with the same presenter, without filming yourself
- 9:16 product picks or livestream ads where prices and numbers need to stand out
- Seeing the generation cost up front and re-shooting one segment at a time
- People already using remote MCP servers and local commands in Claude Code or Codex
Not for
- Generation spends credits on your Renoise account: 482 for the 27-second ad, about 3,250 for the 4:36 tutorial.
- The text-behind-the-presenter effect uses Apple Vision and runs on macOS only.
- Remotion is free for individuals and companies of up to 3 people; larger companies need a company license.
- Swapping Grok's voice for another leaves the lips about 100 ms off. Use only your own face and voice, or ones you have permission for.
Reviewed by Devalley’s editor before publishing. How the column is made