anything2explainer
Topic in, narrated explainer video out — a black-canvas motion-graphics explainer with TTS voiceover, subtitles and a chapter progress bar, in Chinese or English, every frame drawn in code.
Evidence shows four mandatory user confirmation checkpoints (length/language, narration sign-off, voiceover, 30s pilot), clear data-flow disclosure (edge-tts is a Microsoft cloud call, kokoro runs locally), royalty-free B-roll with MANIFEST logging, all frames drawn locally in code, explicit isolation of instructive text inside scraped research docs, no covert actions or external writes. Deducted: parallel builds consume 2–3GB disk with no stated rollback/cleanup path; edge-tts depends on a Microsoft endpoint; license metadata is NOASSERTION so attribution is incomplete. Not full marks.
SKILL.md is highly self-consistent: stages, scripts (tts_build/motion_check/selfcheck/frame_metrics/preview) and checkpoints cross-reference coherently; README corroborates and discloses known pitfalls (pin edge-tts 7.2.8, ARM install issues, pane caps). Deducted: static review finds no test suite or CI execution evidence; script bodies are not in evidence; failure-feedback quality on abnormal input rests on prose only. Capped at 10 by static calibration; scored 9.
Clear trigger semantics (user asks for an explainer video), explicit non-fit ranges (video replication, talking-head, live-action), bilingual support with per-tier length/resource budgets, thorough Linux/macOS environment notes. Deducted: heavyweight environment (Node, ffmpeg, venv, multi-agent parallelism, ≥5GB disk); mainland-China reachability of the Microsoft edge-tts endpoint and kokoro model downloads is not disclosed.
Excellent documentation layering (SKILL.md process + reference specs + lessons + FAQ + known limits), honest acknowledgements and originality statements, full PolyForm Noncommercial license and OFL font licenses. Deducted: repo metadata reads NOASSERTION vs the LICENSE text; no version numbers or changelog; maintenance responsibility and update path are only implicit in a personal repo; validated only on Claude Code/Codex.
The value claim is strong (topic in, finished MP4 with voiceover, subtitles and progress bar out) and the reference film's full paper trail plus contrast frames serve as an auditable yardstick with four cost-controlled checkpoints. Under static review nothing was executed: render quality, QC pass rates and efficiency claims (8 agents / 40 minutes) cannot be independently confirmed, exceeding the static cap of 7 is barred; the 1–3 hour / 2GB cost-benefit is unverified. Scored 6.
Evidence types are relatively rich: a complete reference-film paper trail (research/storyboard/QC reports/frames), 6 contrast frame pairs, two language cuts, and a lessons log distinguishing fact from design intent. Deducted: all material is author-supplied, with no third-party execution or CI+test coverage; the key reproducibility claim (deterministic frame-level rendering) cannot be statically verified. Scored 4.
- This is a static source review only; nothing was executed, so render quality and efficiency claims are not independently verified.
- The default Chinese voice (edge-tts) calls a Microsoft cloud endpoint whose mainland-China reachability is undisclosed; the English kokoro model downloads from overseas sources — verify your network environment.
- Each film consumes ~2–3GB of disk and parallel builds are demanding (≥5GB free, tmux pane caps); no automatic rollback/cleanup path is documented.
- License is PolyForm Noncommercial: free for noncommercial use, commercial use requires prior author authorization; the repo's NOASSERTION metadata conflicts with the LICENSE text.
- Narration is frozen once voiced (hard-coded frame numbers); finalize wording at checkpoint 2.
- Verified only on macOS (plus Raspberry Pi ARM); Windows untested; only Claude Code/Codex have been actually run.
What does this skill do, and when should you use it?
anything2explainer is a Claude Code / Codex skill that turns any topic into a 2–8 minute motion-graphics explainer video, with voiceover, subtitles and a chapter progress bar, all drawn frame-by-frame in code with Remotion (React + TypeScript). No generative video model and no frames lifted from existing footage. What ships here is a whole method, not just a template: a compilable Remotion project, a primitives and lighting library, tooling for voiceover / storyboard / rendering / quantitative QC, written style and motion specs, a multi-agent orchestration protocol, and one complete reference film ("RAG & Knowledge Bases") as the quality bar. A full run takes roughly 1–3 hours of wall clock, mostly parallel agents building shots, with the user consulted at exactly four checkpoints. Bring-your-own TTS is supported, and renders are reproducible (pure functions of frame number with seeded randomness).
Given a topic, length and language, the main session runs nine stages: scaffolds a Remotion 4 project from the template; dispatches a research agent that produces a sourced research doc with a list of numbers and analogies, each with a URL; writes narration following written voiceover-writing principles, runs tts_build.py for the voiceover (default Chinese edge-tts Yunxi, default English kokoro-82m Liam) and gets a frame-accurate timeline and subtitle table; storyboards every shot; adds title / chapter cards / HUD / topic icons; builds the first shot group and renders a 30-second pilot for style sign-off; then dispatches 4–14 parallel build agents, each writing pure-function Remotion components for 5–7 shots with motion_check.py density self-checks; renders a 1280×720 H.264 film and runs frame_metrics.py; finally one QC agent per chapter reviews frames against written criteria, fix agents repair per group, and delivery notes are written. Every number and term on screen must trace to a source URL in the research doc; unverified facts stay off screen and out of the narration.
- A science-communication creator who wants a steady stream of style-consistent explainer videos (e.g. "explain vector databases") without learning animation or editing software
- A technical writer with an existing article or document that should become a narrated, subtitled video for a video platform
- An English-language explainer: the skill has its own English pacing model, subtitle budget and default voice (kokoro Liam)
- An educational content team that wants an agent pipeline to reliably reproduce one visual style, with every on-screen fact traceable to a source
- A developer advocate or marketer who needs a deterministic, code-rendered video — every number and every frame auditable and single-shot fixable — rather than the randomness of generative video models
What are this skill's strengths and limitations?
- Fully code-drawn and reproducible: every frame is a pure function of the frame number, and any bad frame is fixed by editing one shot file
- Strict fact discipline: every number, year and term on screen must trace to a source URL in the research doc; unverified facts stay off screen
- Complete methodology: ships with style specs, a motion vocabulary, six agent prompt templates, a lessons-learned file, and a full reference film as the quality bar
- Well-placed checkpoints put the most expensive rework (script, style changes) at the cheapest point in the pipeline
- No GPU needed: Remotion renders via headless Chromium on CPU; the Chinese voice (edge-tts) is a cloud call and the English voice (kokoro-82m) runs locally on CPU
- Licensed PolyForm Noncommercial 1.0.0: free for noncommercial use, commercial use requires prior authorization from the author (videos you make are yours)
- Landscape 1280×720 only, no 9:16 vertical; one fixed visual style with a two-backdrop switch (stars / dots) — anything else means editing the style guide and ui.tsx by hand
- Narration is frozen once voiced — shot code hard-codes frame numbers, so a rewrite re-times the whole film
- Parallel builds are resource-hungry: several agents bundling Remotion at once need ≥5 GB free disk, and tmux pane caps mean waves of ~4 once past roughly 12 agents
- Only verified with Claude Code and Codex; Windows untested; the GitHub license field reads NOASSERTION (actual license is PolyForm Noncommercial)
- A 3–5 minute film takes ≈2 hours wall clock and ≈2 GB disk, and assumes a harness that can run many agents in parallel reliably
How do you install this skill?
git clone https://github.com/Vincentwei1021/anything2explainer.git, then symlink it for your client: ln -s "$PWD/anything2explainer" ~/.claude/skills/anything2explainer (Claude Code) or ln -s "$PWD/anything2explainer" ~/.codex/skills/anything2explainer (Codex). Dependencies: Node ≥18, ffmpeg (brew install ffmpeg); in a Python venv, pip install 'edge-tts==7.2.8' numpy pillow scipy (edge-tts is pinned because it tracks a Microsoft endpoint and breaks across upgrades); for English narration also pip install kokoro soundfile and install espeak-ng. Scripts are zsh + Python 3, developed and verified on macOS; Linux should work, Windows is untested.
How do you use this skill?
In Claude Code or Codex, just say what you want — e.g. "Make me an explainer video about vector databases" or "讲一下向量数据库，做成一条讲解视频" — and the skill triggers itself. The run stops at exactly four checkpoints: ① length and language (this sets line count, shot count and parallel agent count); ② narration sign-off (once locked, frame numbers are hard-coded into every shot, so changing one word re-times the whole film); ③ voiceover engine (defaults are edge-tts / kokoro, or hand over finished audio at public/assets/<slug>/audio.wav); ④ the first 30 seconds (fixing style here costs one group; after the full render it costs every group). You can also drive the template by hand: template/scripts/new_project.sh ~/work/my-video myslug, then write research/调研.md, script/narration.txt (run tts_build.py), script/storyboard_src.md (run render_storyboard.py), edit src/config.ts, write shot code, preview with scripts/preview.sh 30, and render with VER=v1 scripts/render.sh.
How does this skill compare with similar options?
The README's own positioning: unlike generative video models (Sora, Veo, Runway), output is deterministic code, every on-screen number traces to a source, and any frame is fixable in one file; unlike avatar/presenter tools (HeyGen, Synthesia), there is no presenter — motion-graphics diagrams show the mechanism; unlike hand-rolled Remotion or Motion Canvas, it adds the research → narration → storyboard → parallel build → QC method on top of the canvas; unlike Manim, it is React/TypeScript rather than Python, with TTS-aligned subtitles, chapters and quantitative QC in an agent-driven pipeline.