anything2explainer

A skill for Claude Code and Codex that turns a topic, or a document, into a narrated explainer film. It is not a command-line tool: what ships is the method — a 21,395-byte SKILL.md with nine stages and four checkpoints, ten specification documents under reference/, a compilable Remotion 4 project with a primitives and lighting library, eleven scripts covering the pipeline and its quantitative QC, and one finished reference film with its whole paper trail. Output is 1280×720 H.264 at 30fps in Chinese or English, with chapter cards, a progress bar, a top HUD and subtitles aligned to word boundaries: edge-tts zh-CN-YunxiNeural for Chinese, kokoro-82m am_liam for English, or your own audio. A blank line in the narration is a paragraph and one shot, sentences inside it are 10 frames apart and a paragraph end is 30, so the pause lands where the picture changes and every shot holds 1–1.5 seconds after its last element lands. Every frame is drawn in code, and every number on screen has to trace to a source URL. 2,207 stars, 305 forks, PolyForm Noncommercial.

Who built itThe repository is published under the GitHub account Vincentwei1021, which also publishes video-shotcraft, and the README signs off as Vincent. Its 28 commits all fall between 2026-09-08 and 2026-09-18: 25 of them under this account — signed under three names, Wei Yihao on 17, Vincent Wei on 5 and Yihao on 3, but a single address — one by DHCatLaw, and two more signed mdipaolo1 from an address GitHub links to no account. Twenty-five of the 28 carry a co-author trailer and every one of them names a Claude model: Opus 5 with a million-token context on eighteen, Fable 5.1 on five and Opus 4.8 on two. The contributor list holds two entries: the author with 25 and DHCatLaw with one.

1. 01

   The same author, the same canvas, a different kind of film

   Two records in this archive now come from the account Vincentwei1021. video-shotcraft was created on 2026-07-19 and has 10,029 stars; it exists to make a product look good, and its centre of gravity is a library of 157 shot recipe cards plus a sound-design pass over 149 effects and five music beds. anything2explainer was created fifty-one days later, on 2026-09-08, and has 2,207 stars; it exists to explain a subject, and its centre of gravity is a script. The two share the machinery and almost none of the content: both are skills for Claude Code and Codex, both draw React components through Remotion onto a black canvas, and both are documented at the same length. What differs is what decides the picture. In the card library the shape of the film comes from the shots you pick and the sound you pin to them, and the product is the subject. Here the shape comes from the narration, which is researched, written, voiced and turned into a frame-accurate timeline before a single shot exists — and the shot files then hard-code those frame numbers. There is no sound-effect library here at all, no music bed, no card catalogue and no browser editor. The README draws the line itself: the output is motion-graphics diagrams that show the mechanism, with the narration driving the visuals.

2. 02

   Nine stages, four stops, and a repository that is mostly method

   The entire history is ten days: 28 commits between 2026-09-08 and 2026-09-18, and the last one adds the end credit built by Anything2Explainer skill. Around that sit 2,207 stars, 305 forks, eight watchers, four open items and a 209-file tree. The repository is explicit about what it is not — “It is not a CLI” — and what it carries instead is a process: SKILL.md at 21,395 bytes runs nine stages from scaffolding to QC, and stops at four checkpoints where the run waits for you. Ten documents under reference/ add about 123 KB of specification, led by lessons.md at 32,865 bytes and narration-storyboard.md at 21,950, down to research-brief.md at 3,477. Under template/ sits the compilable Remotion project: src/ui.tsx is 40,704 bytes of primitives and palette, src/fx.tsx 17,209 bytes of light, depth and camera, and eleven scripts of 60 KB whose centre is tts_build.py at 29,185. The heaviest file in the tree is not code: NotoSansSC.ttf, one of four bundled fonts, is 17,772,300 bytes, and the shell scripts are zsh plus Python 3, verified on macOS, with Windows untested. The reception includes one third-party test rather than only stars: the AutoClaw agent workspace reports installing the skill and completing an English explainer task end to end, with the reviewed shots meeting its stated visual criteria.

3. 03

   The narration is the timeline, and a paragraph is a shot

   Pacing here is not a matter of taste; it falls out of the voice track. A blank line in the narration marks a paragraph, and a paragraph is one shot, so the picture changes where the speaker pauses; sentences inside a paragraph are 10 frames apart and a paragraph end is 30. Subtitles are aligned to word boundaries, and the film is allowed to run 5–8% longer than the raw speech so that every shot can hold for 1–1.5 seconds after its last element lands. The second checkpoint is where this becomes expensive: once the narration is voiced, frame numbers are hard-coded into every shot file, so changing one word re-times the whole film. The rule set was rewritten once from real complaints. Viewers of the fifth film said the pacing was too fast and that the purple light sweep kept arriving, and the numbers agreed: of 47 sentences, nine ran under three seconds, and of 40 shots examined in the finished English cut, 26 left no stable frames at all at the end while only seven held for 30 frames or more, and 34 of the 47 shots pushed in. The fix, decided by the user: static time capped at 3 seconds instead of 30 frames, the last beat of a shot holding 30–45 frames before exit, shots at least 120 frames long so that a short sentence becomes a beat inside a neighbouring shot, sentence gaps raised from 10 frames to 20, and at most one highlight moment per chapter.

4. 04

   Where the word boundaries went

   The sharpest piece of forensics in the repository arrived from a user. On 2026-09-09 DHCatLaw filed three problems found while running a four-and-a-half-minute Chinese film on macOS, two of which were real. One was trivial and total: test_render.sh created its temporary directory with a name ending in a dot, and Remotion refuses an image-sequence output directory with an extension. The second looked like a change at Microsoft: across four voices and both languages, WordBoundary events came back as zero, leaving only SentenceBoundary, and pinning edge-tts at 7.2.8 changed nothing. The author reproduced all of it and found the cause on the client side instead: version 7.2.0, released on 2025-08-05, added a boundary parameter to Communicate with a default of "SentenceBoundary", so the request went out saying wordBoundaryEnabled:"false", where 7.0.2 and earlier had hard-coded it true. The failure was silent rather than loud: with no words, subtitle block starts fell back to interpolating by character count. On the repository’s own sample narration, read by zh-CN-YunxiNeural at plus eight percent, the interpolated block starts were frames 0, 9, 45, 81 and 134 against real speech positions of 0, 6, 30, 76 and 130. The fix asks for word boundaries explicitly, and the contributor’s retry for empty audio was adopted with his authorship kept.

5. 05

   What was merged, what was generalised, and what was turned down

   The pull-request queue decided the shape of the project. A Linux and Raspberry Pi port merged after the contributor ran it all on a Pi 5: the default English engine will not install on ARM and Python 3.13, so two local engines were added behind TTS_ENGINE, and REMOTION_BROWSER_EXECUTABLE points Remotion at the system Chromium. A follow-up then audited the skill for generality after finding the sample film written into the method — a fixed chapter formula that treated one subject’s shape as the structure of every film, a shot-pattern table of domain concepts rather than visual relations, and a pronunciation table of one film’s CUDA and NIXL terms — and all three came out. Windows arrived as a report rather than a patch: the pipeline does work, verified on Windows 11, but five shell scripts need replacing, and an open pull request now makes them portable. Two contributions were declined rather than ignored — a spoken title sentence at the head of the film, because the maintainer prefers the existing title shot and timeline, and Russian narration with its tests, as maintenance the main repository would have to carry. The question this record would most like answered — why Remotion rather than another programmable canvas, whether HyperFrames was evaluated, whether the render layer could become a swappable backend — is issue 14, filed with an experiment and still unanswered.

6. 06

   One film shipped whole, and a licence that collects requests

   The reference film is delivered with its entire paper trail, which is more than half the tree: 119 of the 209 files under examples/rag/ — a 48,851-byte research document with a URL for every item, a 35,589-byte storyboard source, nine shot groups holding 44 shot components with a build note of 12,401 to 15,707 bytes each, and five QC reports. That cut is the four-minute-35 Chinese version on a star-field backdrop, built by eight agents working in parallel for forty minutes through two QC rounds; its successor runs 4:54 on the dot-field backdrop, and the English cut runs 5:02 over 44 lines and 785 words voiced by kokoro. The visual language is credited rather than claimed: black canvas, white line art, purple accents and ultra-bold headlines were learned from a Douyin creator named in the acknowledgements, and none of that creator’s assets are used. Originality is stated as two rules: every frame drawn in code, and every fact sourced, with unverified material kept off the screen and out of the narration. Commercial use had to be answered three times: a small United States law firm, an education account publishing to Douyin, WeChat Channels, Xiaohongshu and WeChat, and a Turkish YouTube channel with paid Udemy courses. All three were sent to the author’s address with a request for scope and scale, and the README is clear that the videos made with it belong to whoever made them.