The public GitHub project lanshu-create-ai-presenter-video packages a Codex-oriented Skill for turning a topic or script plus an authorized, clearly adult presenter image into an AI presenter video workflow. Developers should care because the project deliberately does not bind its source code to a particular video, speech or lip-sync provider; instead, it selects from capabilities available in the current environment. The repository describes the resulting process as covering script work, voice generation, presenter generation, lip-sync calibration, captions, keyword motion effects, editing, rendering and quality acceptance.
A workflow rather than a video API
The Skill can be installed under ~/.codex/skills/lanshu-create-ai-presenter-video, and its repository includes job initialization, preflight and delivery-finalization scripts alongside reference documents for generation, editing, and QA recovery. The project page lists Codex or another local Skill-compatible agent environment, Python 3.9+, FFmpeg/ffprobe, standard shell tools, and at least one callable capability each for video generation, speech generation and lip sync as runtime requirements.
Its central production rule is that the completed narration becomes the timeline reference: presenter video, subtitles, shots, keyword treatments and transitions are positioned against the same audio track. The repository documentation says this is intended to reduce lip-sync drift and discontinuities between segments. The documented defaults include 9:16 output at 1080×1920 and 30fps, with topic-led videos generally targeted at 45–75 seconds. The same project page also specifies a roughly -16 LUFS publishing loudness target.
Consent and cost checks are part of the job
The supplied job setup accepts a --rights-confirmed flag and an --adult-presenter-confirmed flag, after which the user is expected to review job.json for manual checks and permission for remote uploads before running preflight validation. The repository states that image-use rights and adult status should be confirmed before remote upload, and that voice-cloning authorization should be confirmed before cloning a voice.
The project also requires an explanation of upload content, generation duration, pricing basis, trial plan and retry limit before the first paid generation. Its documented cost boundary tells the workflow to query existing task IDs after an interruption to avoid duplicate charges, and to stop and summarize the problem after three consecutive paid candidate failures.
The repository says it does not retain API keys, access tokens, signed download URLs or user media, and that task-level request records should have credentials and temporary URLs removed before they are committed. The project page further says that preflight and delivery reports retain filenames rather than absolute paths on the developer's machine. It is published under the MIT License. The repository license information permits use, modification and distribution under that license.
The unresolved boundary of “verified”
Analysis: This is a useful attempt to make consent, spend control and technical QA explicit in an agent-driven media pipeline. But its “verified” outcome is only as reliable as the operator's rights attestations, the selected providers and the checks actually available in the local environment. The project can require confirmations and prescribe lip-sync, presenter, voice and visual checks; the repository does not claim to independently establish image ownership, voice authorization or the truthfulness of a finished video. Provider neutrality therefore improves portability, while leaving the hardest accountability question with the team running the workflow.
