Skip to main content
← Skills
Skill · Code repo

Digital-Human Videos, From Script to Final Edit

Turn a brief into a finished digital-human video. Digital Human Koubo recommends an editing style, generates talking clips through Renoise’s image and video generation service, then handles voice cleanup, timed captions and Remotion editing.

@berryxia Digital Human Koubo
  • Talking-head videos
  • Video editing
  • Digital human
  • Captions

Works with Claude Code · Codex · OpenClaw

Requirements

  • Renoise MCP
  • ffmpeg
  • Python 3 + numpy + Pillow
  • Node.js >= 18
  • whisper-cli

Overview

What you get

  • A reusable presenter profile and a script divided into short talking clips.
  • An editing workflow for explainers, caption-led vertical videos, product launches, tutorials or multiple presenters.
  • Voice cleanup, timed captions, visual packaging and a rendered video.

Good fit

  • Producing recurring product introductions, tutorials or explainers with a consistent presenter.
  • Using an agent to coordinate generation and local editing, with a cost estimate and first-clip review before the full batch.

Tutorial

Understand how the finished video is made

A talking avatar still needs audio cleanup, captions, supporting footage and an edit. Digital Human Koubo joins these tasks in one agent workflow: choose a style for the content, generate talking clips through Renoise, then finish the video with local scripts and Remotion. Renoise is the image and video generation service; MCP connects your coding tool to it.

The video above is the author’s roughly 5-minute-55-second tutorial. Watch for:

  • Presenter reuse: the same character appears in different settings, with additional presenters and animal characters also demonstrated.
  • Speech generation: the tutorial explains generating speech and video together and driving lip sync from an existing recording.
  • Editing for explanation: the presenter shares the screen with steps, demonstrations and emphasized words.

Choose by the result you need

  • Digital Human Koubo takes a brief through digital-human generation, captions and editing to a complete talking-head video.
  • VoiceStudio generates audio locally. It suits an existing video that needs a voiceover.
  • eian-riso-broll makes silent animated inserts from numbers, comparisons and steps in a script. It adds visuals to an existing narration.

Install the skills and connect Renoise

This exercise uses Claude Code and a Bash terminal on macOS or Linux to make a roughly 45-second vertical product introduction. The skill code is free; image and video generation use your Renoise credits.

  1. Run the commands in Quick start from your terminal. They copy both the main skill and its companion, finance-explainer, including their scripts, references and templates.
  2. Enter these two commands in Claude Code, one at a time, to install the Renoise plugin. Follow the browser sign-in when it is first used:
/plugin marketplace add ArcoCodes/renoise-plugins-official
/plugin install renoise@renoise-plugins-official
  1. Start a new conversation and ask it to check your Renoise balance. Once it returns the account credits, check the local tools from your terminal:
bash ~/.claude/skills/digital-human-koubo/scripts/doctor.sh

The check covers FFmpeg, Python with NumPy/Pillow, Node.js 18+, whisper-cli and its speech model. Install the missing items listed in the output, then rerun the check until it reports that the environment is ready. The editing project’s setup script installs its Remotion dependencies.

If you use Codex

From the cloned repository, copy both skills to the Codex directory specified by the author, then connect the MCP server:

mkdir -p ~/.codex/skills
cp -R skills/digital-human-koubo skills/finance-explainer ~/.codex/skills/
codex mcp add renoise --url https://renoise.ai/api/public/mcp
codex mcp login renoise

Start a new conversation and explicitly ask to use digital-human-koubo. Replace ~/.claude/skills/ with ~/.codex/skills/ in the terminal commands below. The author made the examples with Claude Code; the Codex setup follows the repository’s compatibility instructions.

Prepare your presenter and content

  1. Create a working folder. In brief.md, record the product name, audience, two verified features, material to show and the action viewers should take.
  2. Put a clear, evenly lit photograph of yourself at assets/portrait.jpg. Add product photographs or screen recordings to assets/.
  3. Open this working folder in Claude Code and send this tutorial exercise:
Use digital-human-koubo to make a roughly 45-second, 9:16 product introduction.
Read brief.md. Create my digital human from assets/portrait.jpg and use the product material in assets/.
Recommend two editing styles with preview images and wait for my choice. Keep captions easy to read.
Then prepare the script and presenter profile. List the images to generate, talking-clip count, duration, estimated Renoise credits and my balance.
Wait for my cost approval before each generation. Make only the first segment for review before generating the full batch.

The skill recommends two options from these five styles:

LabelStyle
AExplainer
BVertical video with animated captions
CLaunch breakdown
DStep-by-step tutorial
EMultiple presenters or languages

For this exercise, choose Style B, the vertical caption-led format. It suits a short product introduction and uses edit.json to control captions and visual elements. If B is not among the recommendations, ask to use it instead.

For a new presenter, the skill prepares reference views and test clips, then saves the selected appearance and voice description in persona.json. brief.json holds the style and aspect ratio; script.json holds the segmented narration. The brief guides the edit, while the script determines what is said.

Review the first clip before generating the rest

  1. Read the script. Correct product facts, number pronunciation and the closing action. Keep every segment within the model’s 15-second limit.
  2. Check the quote for images, clip count, seconds and credits. Once you approve it, have the agent generate only the first clip.
  3. Open that clip. Listen to the voice and inspect lip sync, appearance and framing. Resolve these choices before approving the remaining clips.

When the first clip works, continue with:

The first clip's appearance and voice are approved. Keep the same presenter profile and make the remaining clips covered by the quote I approved.
Transcribe and compare each clip with the script, inspect frames, then make the audio consistent across clips and create word-timed captions before editing.
Use Style B with restrained captions. Show the corresponding product material for each feature and retain the closing action from brief.md.
If you need extra generations or retakes, explain why and quote the additional credits before proceeding.

This order resolves the choices that affect every clip early. Changing the presenter or voice affects the whole production; changing a caption color or card position usually affects only the edit.

Finish the edit and inspect the result

The skill downloads and checks clips, uses Whisper to compare speech with the script and build timing data, then processes the audio and edit. In a Style B project, these files have separate roles:

FilePurpose
public/main.mp4The assembled digital-human footage
public/voice.wavThe processed voice track
public/show/Product material to display
src/tl.jsonWord and phrase timing
src/edit.jsonCaption intensity, keywords, cards, product inserts and the closing action

Ask the agent for the project path and a preview. Check that captions are readable, product inserts match the narration and music leaves the voice clear. Inspect frames during transitions for overlapping elements. Compare captions with what you actually hear.

For example, if the inserts move too quickly, give a focused revision request:

Keep the generated talking clips, voice track and total duration. Change only the edit: hold the first product image for at least 3 seconds, remove the emoji beside the face and retain the current keyword accent color.
Show me a preview of this section before exporting the complete video.

Before export, the agent can check timing references with the bundled checker. To run it from a terminal, replace ./product-edit with the actual editing-project path:

python3 ~/.claude/skills/digital-human-koubo/scripts/check_edit.py ./product-edit/src

The message ✓ 时间引用全部能对上 means that timing references passed the check. If it reports ✗ errors, ask the agent to fix them and rerun the check. Once the check passes and you approve the preview, ask the agent to export the complete video and return the final MP4 path. Open that file to confirm that captions, product inserts and audio are included.

The skill’s code and documentation use the MIT license. Retain the video’s AI-presenter label. If you use the bundled animated emoji, keep the credit Animated emoji: Google Noto in the ending. See the third-party notice for Remotion’s team licensing terms.

Make your next video

Reuse the same persona.json for a recurring presenter. For another product, replace the script and demonstration material. For a horizontal tutorial, reconsider the structure and style. Styles C, D and E provide reference editing implementations whose scenes need adapting to the new script. Style A uses the companion finance-explainer engine installed earlier.

Replace the brackets with your own details:

Use digital-human-koubo to make a [product introduction / tutorial / launch breakdown].
Source material: [script, photograph, screen-recording or product-image paths].
Audience and action: [who will watch and what they should do afterward].
Presenter: [existing persona.json path, or a photograph for a new presenter].
Platform, aspect ratio and target duration: [your requirements].
Recommend an editing style first, then confirm the script and cost. Wait for my first-clip approval before continuing.
Editing priorities: [caption intensity, material to show, numbers or steps to emphasize].

The optional text-behind-presenter effect needs macOS person segmentation. Request a medium shot with background space for the text before generating that clip. Complete the basic workflow before adding this effect.