Digital Human Koubo Skill: One Photo to a Finished Talking-Head Video

The Digital Human Koubo Skill helps educators, product marketers and video creators turn an authorized presenter photo and a script into a packaged talking-head video. It does more than animate a face: the workflow chooses an editing style, splits the script into short generation jobs, estimates paid credits before spending them, checks the spoken output, unifies the voice, and assembles the final video in Remotion.

Skill Summary

The Digital Human Koubo repository contains two installable Skills, but they are not two equal products. digital-human-koubo is the main production orchestrator. finance-explainer is a companion engine used when the main workflow selects Style A, the longer finance or technology explainer format.

• Input: an authorized presenter image or persona, a script or topic, and choices such as aspect ratio, duration and caption intensity.
• Generation: short lip-synced clips created through Renoise MCP, with a cost estimate before any paid batch starts.
• Editing: one of five production styles, plus a repository of packaging components for hooks, evidence cards, maps, chapter cards, captions and calls to action.
• QA: speech-to-script comparison, frame inspection, voice matching, transition checks and loudness normalization.
• Output: a rendered talking-head video and a local Remotion project that can be revised.

Where This Skill Fits

This Skill fits between a simple avatar generator and a full manual video-editing project. A basic generator may return a presenter clip, but it does not necessarily decide how to pace a three-minute explanation, where to place supporting evidence, or how to keep separately generated clips sounding consistent. Digital Human Koubo treats those decisions as one repeatable production pipeline.

It is best suited to teams that already know what they want to say and have permission to use the presenter identity. It is not a substitute for factual research, product approval, image rights or a final human review of the voice and likeness.

Who It Helps

• Educators and course teams: turn a lesson script into a tutorial with chapter cards, screenshots and a consistent presenter.
• Product and marketing teams: produce a short launch explanation or vertical product-recommendation video while keeping a review gate before paid generation.
• Finance and technology creators: use the companion finance-explainer engine for sourced facts, charts, documents and longer narrative structures.
• Multilingual publishers: build multi-presenter or multi-language episodes, then verify each generated segment instead of assuming the speech is correct.

Access / Install Links

The full pipeline needs three capabilities: an agent that can load Skills, a Renoise MCP connection for image and video generation, and permission to run local commands for downloading, checking and editing files. The repository documents Claude Code as tested and lists Codex, OpenClaw, WorkBuddy and other Skill-capable agents as possible hosts when those three requirements are available.

• Local tools: FFmpeg, Python 3 with NumPy and Pillow, Node.js 18 or later, and whisper.cpp with a model for transcription checks.
• External service: a Renoise account and MCP connection. Generation consumes credits, so current model availability and pricing should be checked before production.
• Your own references: the material: identifiers in the repository belong to the author and will not work in another account. Create and register your own authorized persona reference.
• Optional voice service: Xiaomi MiMo TTS can provide a designed or permitted cloned voice, but it introduces a separate account, API key and pricing boundary.

Setup Steps

Read the current repository instructions before copying files. The commands below show the documented structure; adjust the destination for your agent rather than installing blindly into an unrelated environment.

1. Review the repository and its current commit before installing executable instructions.
2. Copy both Skill folders into the directory used by your agent.
3. Connect Renoise MCP, sign in through the browser and confirm that the agent can read the account balance.
4. Run the repository’s doctor script before local editing, then install only the missing FFmpeg, Python, Node or Whisper dependencies.
5. Replace the author’s unusable reference IDs with a presenter image and voice you are authorized to use.

Example Inputs

git clone https://github.com/andyhuo520/digital-human-koubo.git
cd digital-human-koubo

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

How to Use

1. Check the environment: confirm Renoise access and run the repository’s doctor script only when local editing is required.
2. Diagnose the content: the Skill recommends two of five editing styles and asks for the style, ratio, duration, caption intensity and presenter in one round.
3. Write and segment the script: each spoken segment is designed to stay within the model’s short-clip limit, normally around ten seconds and never longer than 15 seconds.
4. Quote before spending: the Skill reports segment count, duration, estimated credits and account balance, then waits for explicit approval.
5. Test one segment: it generates the first clip so the user can judge the face, voice and scene before the remaining batch consumes more credits.
6. Download and verify: the scripts use range downloads, decode checks, Whisper transcription comparison and four sampled frames per segment.
7. Unify and align audio: spectral matching, de-essing, shared room tone and loudness normalization reduce the audible difference between separately generated clips.
8. Edit in Remotion: the selected template combines the presenter, captions, evidence and packaging components, then renders the final video for transition and loudness checks.

Common Mistakes & Fixes

• Batching before testing: approve the quoted cost, generate segment one, and inspect the likeness and voice before creating the rest.
• Reusing the author’s material ID: upload your own authorized reference and replace the account-specific ID in persona.json.
• Trusting transcription alone: also watch the sampled frames and the complete render for burned-in captions, visual drift and bad transitions.

Selection Logic

A · Finance or Tech Explainer

This is where the companion finance-explainer Skill enters. It adds a sourced fact sheet, narrative structures, charts, document callouts and a longer horizontal-video engine. It should learn the structure of successful explainers without copying a real creator’s face, voice, set or wording.

B · Vertical Caption-Led Talking Head

A 9:16 format for product recommendations, short lessons and sales-oriented clips. The template supports three caption-intensity levels and faster visual events, but the content still needs truthful product claims and authorized assets.

C · Launch Breakdown

A restrained 16:9 format for explaining a release, model or product. The presenter can move to one side while source material, interface footage or code appears beside them.

D · Step-by-Step Tutorial

A longer horizontal format that combines the presenter with screen recordings, close-ups and chapter structure. It is designed for instructional pacing rather than a dense short-form edit.

E · Multiple Presenters or Languages

A format for alternating hosts, scenes or languages. The repository reports tests in several languages, but every new language and voice should still be transcribed and reviewed before publication.

Expected Outputs

• A reviewed script and brief: style, ratio, duration, caption intensity and presenter choices are written down before generation.
• Verified short clips: each segment has a corresponding task, downloaded media, speech comparison and sampled-frame check.
• A local editing project: Remotion source, aligned timing data, captions, audio and selected packaging components remain available for revision.
• A publishable render candidate: the final file has consistent voice treatment and target loudness, but still requires a complete human watch-through.

Use Cases

• Course tutorials: combine a presenter with screen recordings, chapters and explanatory cards.
• Product launch explainers: place the presenter beside source material, interface footage or code.
• Vertical recommendations: create a caption-led 9:16 edit after checking every product claim.
• Long-form finance or technology stories: use the companion engine to connect sourced facts, charts, documents and narration.

Why the QA Gates Matter

The strongest design choice is not a particular avatar model. It is the sequence of gates around an unreliable and paid generation step. The Skill asks for approval before spending credits, validates one segment before batching, checks whether the generated speech matches the script, samples frames for accidental burned-in subtitles, and separates the spoken audio from the edit so multiple clips can be normalized together.

Those checks cannot promise a publishable result. They make failures easier to find while the project is still cheap to repair. A person still needs to watch the complete video at normal speed and verify likeness, pronunciation, factual accuracy, transitions and sound.

Limitations

• Use only authorized identities: create a digital human from your own photo or one you have explicit permission to use. Do not imitate a real creator’s face or cloned voice.
• Label synthetic media: the repository requires the finished video to disclose that an AI digital human performs the content, with additional labels for generated scenes.
• Paid generation remains external: Renoise models, credit rates and availability can change. Re-check the quote before each production batch.
• Some effects are platform-specific: text placed behind the presenter uses Apple Vision person segmentation and therefore needs macOS.
• Licenses have layers: the repository is MIT, while Remotion, fonts, voice services, generation models and any music or footage keep their own terms. The repository notes a separate Remotion company-license boundary for larger organizations; verify the current terms directly.
• This is a new repository: inspect updates, scripts and external dependencies before using it in a production account. Big Prompt Hub reviewed the documentation at a fixed commit but did not spend credits or benchmark an end-to-end render.

Related Tools / Prompts

Renoise MCP supplies the paid image and video generation layer; Remotion supplies the editable composition and rendering layer; FFmpeg and whisper.cpp support local media checks. None of them replaces the Skill’s orchestration rules, and the Skill does not replace each service’s current documentation, pricing or license terms.

FAQ

• Q: How do I install the Digital Human Koubo Skill?
  A: Clone the repository, review its scripts, copy both Skill folders into your agent’s Skill directory, connect Renoise MCP, and run the local dependency check before editing.
• Q: Why does the repository include two Skills?
  A: digital-human-koubo controls the overall production flow. finance-explainer supplies the research and editing engine for Style A, so it is a companion rather than a second equal product.
• Q: Can the Skill create a digital human from any photo?
  A: The workflow starts from a reference image, but you should only use a photo and voice you own or are authorized to use. The included account-specific reference IDs cannot be reused.
• Q: Does it generate the whole video in one request?
  A: No. It plans the format, splits the script, quotes credits, tests one segment, generates the remaining clips, checks them, unifies the audio and then edits the result.
• Q: Can I use it without local command access?
  A: A chat-and-MCP client can help plan, quote and generate clips, but the documented download, Whisper QA and Remotion editing workflow needs a shell-capable environment.

Install or adapt this skill for an authorized talking-head workflow? Share what you build in the comments.

Explore more reusable capabilities in AI Skills or more production workflows in Video & Music.

Share your stack or follow @bigprompt for more.

Related Big Prompt Hub Pages

For an editable code-animation workflow rather than an AI presenter, see huashu-motion Skill：让编码 Agent 制作可编辑的产品发布片. If you need a direct video-generation prompt, compare the lighter Veo 3.1 Video Ad Prompt: Miniature Presenter Product Demo.

Teams working on discoverability can connect the finished video workflow to GEOInflow AI Crawler Analytics: Self-Hosted Website Evidence. For a visual prompt outside the talking-head category, explore Travel Sketchbook Illustration Prompt: Hand-Drawn City Journal, or review GPT Image 1 Migration: Move API Workflows to GPT Image 2.5 before updating an image pipeline.

Big Prompt Hub Review

Digital Human Koubo is most useful when the real problem is not “make a face talk,” but “turn many short, fallible generations into one reviewable video.” Its cost gate, first-segment checkpoint and QA chain are more important than the promise of one-photo automation. Start with a short authorized test, verify the voice and likeness, and only then spend credits on a longer production.