A talking-head video pipeline that agents can't fake their way through

What it does It’s an Agent Skill — a SKILL.md playbook plus a few Python and Bash scripts — that steers a coding agent through the entire production of a short talking-presenter video: script, narration, presenter generation, lip-sync, captions, keyword motion graphics, editing, rendering, and QA. You supply a topic or script and one authorized portrait; you get back a master, a share encode, a contact sheet, and a delivery report. It loads in Claude Code, Codex, Gemini CLI, Cursor, OpenCode, GitHub Copilot — anything that can read files and run shell commands.

The interesting bit Progress is a state machine computed from evidence, not declared by the agent. check_state.py derives one of eight states, intake through verified, from the job’s artifacts and exits non-zero if the recorded state overclaims — so “done” means a decodable render, an ASR report, and passing loudness, not the agent’s word. Money gets similar care: a billing statement before the first paid call, retry ceilings, and recovery by saved task ID so an interrupted job isn’t resubmitted and billed twice.

Key highlights

• Harness-neutral: one skill file works across Claude Code, Codex, Gemini CLI, Cursor, OpenCode, and other Agent Skills clients.
• Provider-neutral: voice, presenter video, lip-sync, and speech recognition are selected by capability at run time; each job records the provider, model, parameters, and task IDs actually used.
• Audio-locked timeline: the approved narration is the master clock for presenter motion, captions, cuts, and final duration.
• Verified delivery: master and share encodes are fully decoded and loudness-checked (−16 LUFS ±0.5 LU, true peak ≤ −1 dBTP) before anything is published.
• A smoke test drives a synthetic job from intake to verified without calling any remote service.

Caveats

• It’s choreography, not generation: you must bring at least one voice synthesis, presenter video, and lip-sync capability — a cloud CLI, API, or local model. The repo ships none of them.
• The toolchain is Unix-flavored: Bash, FFmpeg with ffprobe, jq, awk, sed.
• The deterministic compositor for captions and motion graphics (e.g., HyperFrames) is optional, and the README doesn’t say what you get without it.

Verdict Worth a look if you already live in an agent harness and want to delegate short-form presenter videos without babysitting — or if you want a template for evidence-gated agent workflows in general. Skip it if you expected a self-contained generator; the skill is only as good as the providers you point it at.