anything2explainer packages a Claude Code and Codex skill that turns a topic, article or document into a narrated explainer video. Developers get a code-first production path: every frame is drawn with Remotion in React and TypeScript, rather than assembled from stock footage or produced by a generative video model, so individual shots remain editable as source code.
The project targets a 1280×720, 30fps H.264 MP4 with synchronized voiceover, word-aligned subtitles, chapter cards, a top HUD and a chapter progress bar. It supports Chinese and English and delivers the research document, narration, storyboard, per-shot source code and QC reports alongside the finished video, according to the repository specification.
A production method, not a CLI
The repository explicitly describes the project as something other than a CLI. It ships a compilable Remotion template, visual primitives, lighting components, voiceover and rendering tools, style and motion specifications, a multi-agent work protocol and a complete reference film intended to establish the quality target for a run, as detailed in the project documentation.
Its prescribed workflow has nine stages: scaffold the project; research the subject; write the narration and frame-accurate timeline; storyboard each shot; build overlays and topic-specific primitives; produce a pilot group and 30-second preview; construct the remaining shots in parallel; render the full film and collect frame metrics; then run QC, apply fixes, re-verify the result and prepare delivery notes. Build agents receive groups of five to seven shots and write Remotion components for them, according to the documented process.
Narration structure controls the edit. A blank line marks both a paragraph and a shot, while TTS word boundaries feed the subtitle table and frame-level timeline. The repository says shots retain a short hold after their final element appears, making the completed film longer than the raw speech by design under its timing rules.
The process stops at four checkpoints: selection of length and language, narration approval, choice of voiceover engine and review of the first 30 seconds. The narration checkpoint occurs before voice generation because shot frame numbers are subsequently fixed to that timeline, while the preview is rendered before all shot groups are built to allow an earlier style decision.
The repository estimates one to three hours of wall-clock time depending on video length. Its reference tier for a three-to-five-minute film specifies 40–50 shots, eight build agents and roughly two hours; these figures are the project’s planning estimates rather than an external benchmark published by the maintainer.
Manual control remains available
For agent use, the documented installation symlinks the repository into either the Claude Code or Codex skills directory. That setup is optional: developers can instead invoke the project template, TTS, storyboard, preview, render and frame-metrics scripts by hand using the documented sequence.
The listed environment includes Node 18 or later, FFmpeg and Python packages selected partly by the TTS engine. The scripts were developed and verified on macOS; the repository says Linux should work, documents a verified Raspberry Pi 5 configuration and labels Windows untested in its platform notes.
Analysis: inspectability versus pipeline ownership
Analysis: The project’s defensible distinction is not automatic video creation by itself, but the conversion of production decisions into inspectable artifacts: sourced research, narration, timing tables, storyboards and shot components. That structure may make a factual correction or visual revision easier to isolate than it would be in a single generated clip. The same design leaves the developer responsible for dependencies, TTS selection, agent execution and final review, all of which remain explicit parts of the documented workflow. The unresolved trade-off is whether that added control justifies operating a multi-stage video toolchain rather than using a less editable service.
