Image: Digital Human Koubo
Why it mattersA small team can draft a talking-head ad on their own face before paying a production shop. The skill reports the credit cost and asks for approval before any video model call.
A small team that wants to test a product ad or a founder update usually faces a choice. Book a studio day the team cannot afford. Or buy a stock-footage video that does not say the thing. Digital Human Koubo, a new open-source agent skill by Berryxia, opens a third door. Give it one clear photo of a face and a line about what the video should say. The skill writes the script. It cuts the script into clips. Each clip shows a lip-synced version of that face. The skill cleans the voice, burns the captions, and makes a finished MP4. The repository is MIT licensed. It has 1,866 stars added in the past 30 days.
What the skill makes
The author lists five editing styles in the README. Style A is a finance or tech explainer at 16:9, three to twenty minutes. Style B is an influencer talking head at 9:16, 30 to 120 seconds. Style C is a product launch at 16:9, 30 to 90 seconds. Style D is a step-by-step tutorial at 16:9, three to eight minutes. Style E is a multi-host piece at 16:9, three to seven minutes. The agent reads the user's one-line request. It picks two of these styles as a suggestion. Then it waits for the user to choose one.
What it costs to run
The skill calls Grok Video 1.5 through the Renoise MCP server. Berryxia writes the per-second cost on the README. Grok Video 1.5 at 720p is 17 Renoise credits a second. At 480p it is 10 credits a second. GPT Image 2.5 for the cover is 23 credits for a 2K image. A 27-second livestream ad in the demo cost 482 credits. A 4 minute 36 second tutorial cost about 3,250 credits. A 5 minute 54 second follow-up cost about 3,409 credits. The skill quotes the price to the user and asks for approval before it starts.
Where it checks itself
Every clip is sent back through whisper and compared with the written script. The README puts the match at 94 to 100 percent. Four frames per clip are looked at by eye to catch a subtitle that was burned into the video. Voices are brought to one sound through one-third-octave spectral matching, de-essing, a single shared small-room reverb, and loudness normalisation. The face, the clothes and the voice are kept the same across clips through a shared persona.json file.
What the author refuses
Every finished video carries the line "performed by an AI digital human". The README says users must only use their own face and voice, or ask the owner first. Made-up numbers, prices, offers, or fake reviews are not allowed in the script. The skill works with Claude Code, Codex, OpenClaw and WorkBuddy. Any agent that can load an Agent Skill, call a remote MCP server and run local commands will work.
A short talking-head video used to need a studio day, a camera, and an editor. A team already paying for a video model can draft that video for a few hundred credits. Commit to the real production only if the draft lands.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.