make-explainer-animated-videos
Turn a voice-over recording plus its script into a premium dark-mode animated explainer video for Reels / Shorts / TikTok (1080x1920): particle hero that morphs between shapes, thin wireframe 3D diagrams with glowing active parts and travelling packets, word-synced captions with underlined key terms, progress rail, rolling counters, stamps, recap list, comment-CTA bubble, plus cleaned voice (silence, filler and breath trimming), a properly mixed music bed, role-based SFX, camera punches and transitions, frame-0 headline, covers A/B, subtitles, posting kit and a gated upload. 100% free and open source (HyperFrames + Canvas + ffmpeg + whisper.cpp), no footage, no AI video, no paid tools. Use when the user says "/make-explainer-animated-videos", "make an animated explainer video", "animated explainer from this voiceover", or hands over a voice file + script and wants this kind of animated infographic video.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/avijeetpandey--make-explainer-animated-videos)HTML
<a href="/listings/avijeetpandey--make-explainer-animated-videos"><img src="/api/listings/avijeetpandey--make-explainer-animated-videos/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
# as a Claude Code skill git clone https://github.com/avijeetpandey/make-explainer-animated-videos ~/.claude/skills/make-explainer-animated-videos bash ~/.claude/skills/make-explainer-animated-videos/scripts/setup.sh # checks tools, downloads the whisper model, GSAP and fonts
Standalone (no Claude): clone anywhere and run the scripts in `scripts/` (see `docs/USAGE.md`).
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 57 lines
- ---
- name: make-explainer-animated-videos
- description: Turn a voice-over recording plus its script into a premium dark-mode animated explainer video for Reels / Shorts / TikTok (1080x1920): particle hero that morphs between shapes, thin wireframe 3D diagrams with glowing active parts and travelling packets, word-synced captions with underlined key terms, progress rail, rolling counters, stamps, recap list, comment-CTA bubble, plus cleaned voice (silence, filler and breath trimming), a properly mixed music bed, role-based SFX, camera punches and transitions, frame-0 headline, covers A/B, subtitles, posting kit and a gated upload. 100% free and open source (HyperFrames + Canvas + ffmpeg + whisper.cpp), no footage, no AI video, no paid tools. Use when the user says "/make-explainer-animated-videos", "make an animated explainer video", "animated explainer from this voiceover", or hands over a voice file + script and wants this kind of animated infographic video.
- ---
- # make-explainer-animated-videos
- The user records a **voice-over** (English, or Roman-script Hinglish) and has a **script**. This skill cleans the voice, designs the visuals (a storyboard in `scenes.json`), builds a HyperFrames project, masters the audio, checks everything, renders a draft, and after approval produces the final MP4, covers, subtitles and posting kit. Read `references/STYLE_ANALYSIS.md` (what the look is and why) before storyboarding.
- **If `PERSONAL.md` exists in this folder, read it first: it holds the owner's handle, approval rules, paths and channel rules, and overrides the defaults below.** Per-user settings live in `config.local.json` (see `docs/USAGE.md`).
- `K` = this folder (e.g. `~/.claude/skills/make-explainer-animated-videos/`) · projects live in the `projects_dir` from config (default `~/Documents/explainer-videos/<slug>/`) · `examples/sample/` is a working storyboard.
- ## Hard rules
- - **Free + open source only:** HyperFrames (Apache-2.0, HTML to MP4 in headless Chrome), plain Canvas/CSS (no WebGL, no GPU needed), ffmpeg, whisper.cpp (MIT), fonts under SIL OFL, SFX synthesised by `scripts/make_sfx.py`. **GSAP** is used only as the paused timeline clock: free under GreenSock's own licence, not OSI open source (swap for Anime.js if that matters).
- - **The user's voice is the audio, and it comes first.** Never generate a voice (a TTS voice is for smoke-testing the pipeline only). It is cleaned once, up front, by `clean_voice.py` (`references/VOICE_CLEANUP.md`); the original file is never touched; the music is solved around it, never the other way round.
- - **Every video gets a music bed** from a library the user may legally use (`config.local.json` `music_dir`, built with `scripts/build_music_index.py`); the licence line is copied to `assets/music_credit.txt`. Default style `premium` = techno / tech-house / house / technology beds with a dark spectrum and a driving groove, never pop, funk, corporate or ambient (`python3 -I K/scripts/audio_lib.py --rank premium`; `alt_mix.py` A/Bs another track in seconds). Mix target: voice 10-14 dB over music + SFX (median), never closer than 8 dB, master -14 LUFS, true peak <= -1.5 dBTP.
- - **Copy the format, not a brand.** The header shows the user's own handle (`meta.handle` or config). Never reuse another creator's name, characters, scripts or examples.
- - **Never invent facts.** Every counter/bars carries `"src"` or is auto-labelled as a simplified example; do not assert company/product facts the research could not source.
- - **Approvals:** ask whenever a sign-in is needed; show the draft + contact sheet before calling anything final; upload only after the user approves BOTH the video and the covers; never post anything for the user.
- ## Workflow
- 0. **Setup (once per machine):** `bash K/scripts/setup.sh` (Node 22+, ffmpeg, whisper-cli + `ggml-small.en` model, GSAP cache, fonts, numpy). Optional: `pip install librosa` for `build_music_index.py`. The first render downloads Chrome (1-2 min).
- 1. **Inputs.** Need the voice file (wav/mp3/m4a/aiff) and the script. Create `<projects_dir>/<slug>/` with `voice.<ext>` and `script.txt`. Ask for exactly what is missing. Check the voice with `ffprobe` (>= 44.1 kHz, no clipping); 8 kHz phone audio: ask for a re-record.
- 2. **Voice cleanup (automatic inside the build).** `clean_voice.py`: first word ~0.05 s in, sentence gaps capped at 0.22 s and phrase gaps at 0.12 s (cut from the middle, breaths go with it), standalone fillers (um/uh/ahh/hmm) deleted, drawn-out vowels shortened, repeated phrases keep the last take, +/-3 dB level matching across recorded parts, then a voice chain (high-pass, denoise, EQ, de-ess, compress). Writes `voice_clean.wav` + `assets/voice_edit.json` (every removed span with its reason): READ it at the checkpoint. A hand-timed `words_override.json` stays in the ORIGINAL voice's timeline and is re-mapped automatically. Off: `meta.voice_clean: false`.
- 2b. **Language.** English: nothing to set (whisper aligns the script). **Roman-script Hinglish** (or any language whisper cannot read): `"language": "hinglish"` in `meta` = island-anchored timing (acoustic speech islands + per-island whisper-medium with a Roman prompt as soft evidence, fuzzy-aligned to the script); a `words_override.json` always wins. Ask for ~0.3-0.5 s pauses between sentences when recording. Island starts are reliable; words inside a fast island are +/-0.2-0.4 s. The caption font has no Devanagari: keep scripts in Roman letters.
- 3. **Storyboard -> `scenes.json`** (schema + recipes: `references/STAGE_LIBRARY.md`; working example: `examples/sample/scenes.json`):
- - **One spoken sentence = one scene**, 8-18 words; split long sentences at commas so every half gets its own visual. `text` must be the sentence exactly as spoken; mark the 1-3 meaning words with `*key*`.
- - Hook scene (hero already moving, the question, topic word underlined), a concrete everyday example, the mechanism one step per sentence, a failure/twist (red), the fix (green), a recap `list`, then the `cta` bubble.
- - One new visual idea per scene; persist the diagram across scenes (same `id`s, `keep`). Stage content in y 480-1450, cards <= 780 px wide, hero offset -250...-350 above a low diagram, nothing below y 1500.
- - Colour = meaning: one accent per concept (primary/active, second, third, green success, red failure). Land key visuals on spoken words with `"at": "<word>"`. `"shock": true` on the 3-4 reveal scenes.
- - Facts: `"src": "<where the number comes from>"` on counters/bars, or they are labelled "SIMPLIFIED EXAMPLE".
- 4. **Build:** `python3 -I K/scripts/build_explainer.py <proj>` -> `index.html` + `assets/` and it masters the audio (`assets/audio/master.wav`, `assets/mix_report.json`). It prints the world, voice cleanup, facts, transitions/punches, music, mix balance and "script/voice match": below 80% means the script differs from the recording. `python3 -I K/scripts/mix_audio.py <proj>` re-solves only the mix (seconds).
- 5. **Check:** `cd <proj> && npx --yes hyperframes@0.8.209 check` (0 errors; accepted: canvas-edge info, `line_unread` on fast speech, `text_below_floor` on mono texture labels). Fix every contrast/overlap finding.
- 6. **Look:** `npx --yes hyperframes@0.8.209 snapshot --at <scene midpoints>`; also `T-0.07, T, T+0.1` of every transition. Nothing covered, nothing overlapping, the incoming scene already shows content at the cut.
- 7. **Draft:** `npx --yes hyperframes@0.8.209 render -o renders/picture.mp4 -q delivery -f 30` (`-q draft` while iterating), then `bash K/scripts/finalize.sh renders/picture.mp4 renders/<slug>_final.mp4 assets/audio/master.wav`, then `python3 -I K/scripts/qa_explainer.py <proj> renders/<slug>_final.mp4` (resolution, loudness, mix balance, voice cleanup stats, SFX density, silent spots, visible change <= 3.2 s, budgets; contact sheets in `qa/`) and `python3 -I K/scripts/safe_zone_check.py <proj> renders/<slug>_final.mp4`. Max 3 fix rounds.
- 8. **Checkpoint (one message):** the video, `qa/` sheets, a one-line list per scene, the length, the voice-edit list, the facts register, anything not verified. Wait for approval.
- 9. **Publish pack** (all gated; details `references/PUBLISHING.md`): `meta.headline` (frame-0 claim: line 1 the belief, line 2 the twist) -> `make_cover.py` (covers A/B + 3:4 grid check) -> `make_srt.py` -> research (`seo_research.py` + web searches, save `research_notes.md`) -> write `posting_kit.md` from `assets/posting-kit-template.md` -> `kit_lint.py` PASS -> LinkedIn (`linkedin_post.md` + `make_linkedin_image.py`) -> `publish_check.py`.
- 10. **Upload:** `bash K/scripts/upload_drive.sh <proj> --dry-run` shows the plan; the real run needs `--approved-video --approved-cover` after the user approves BOTH, goes to a NEW `<drive_dest>/<reel>/` subfolder (`<reel>_vN` -> `<reel>/vN/`) and verifies sizes. Never otherwise.
- ## What makes it hook (built in; `references/ENERGY_AND_SOUND.md`)
- Opening camera snap, slow drift, camera punches on number/negation/key words (every 3rd with shake), at most one transition per ~9 s at idea changes (whip / focus pull / vertical whip / push, each with its own sound), impact rings on `shock` reveals and stamps, riser -> hit before the first stamp, beat-locked glow and particle pulse, film grain, a gap filler so nothing sits still for > 3 s, music + SFX by role, no empty frames after a cut.
- ## Worlds (variety)
- `assets/worlds.json`: mono_pulse, blueprint, amber_signal, terminal_green (palette, background, caption style, overlay, motion strength, transition order, preferred music genres). `meta.world` or auto-rotation that never repeats either of the last two videos (`<projects_dir>/_style_log.json`); a re-edit keeps its world and track. Add your own: `docs/TWEAKING.md`.
- ## Options (`scenes.json` -> `meta`)
- `world` - `language` ("hinglish") - `handle` - `series` - `headline` - `cover` - `voice_clean` / `voice_chain` / `sentence_gap` / `phrase_gap` / `voice_reclean` - `music` (false) / `music_style` ("premium" | "classic") / `music_skip` / `music_bpm` / `music_file` - `objScale` (1.3) - `drift` - `hinglish_asr` (false = pause-based only) - `illustrative_label` - scene `shock`. Element and hero fields: `references/STAGE_LIBRARY.md`.
- ## How it works
- `build_explainer.py` cleans the voice, aligns script words to time, resolves element lifetimes and keyframes, plans camera/transitions/SFX (`fx_plan.py`), picks music and SFX (`audio_lib.py`), masters the audio (`mix_audio.py`) and inlines everything into one `index.html`. `engine.js` makes **every pixel a pure function of time** (`render(t)`), driven by one paused GSAP tween: canvas = dust + ~4k-particle morphing hero + wireframes/links/packets; DOM = captions, chips, cards, counters. Deterministic and seekable; seeded RNG only. `docs/ARCHITECTURE.md` has the map.
- ## Known limits
- Wireframes are edges (no solid faces); one particle hero at a time; Canvas 2D (raise `meta.particles` to 6000 if the machine copes); the typefaces are the closest free matches to the studied look; no presenter/avatar; automatic filler detection only catches sustained sounds ("aaa", "ummm"); Hinglish word timing inside fast islands is approximate; the mix is judged by numbers, so listen once with headphones before the checkpoint. Render is about 2x real time for a draft, 3x for delivery.
- ## Study mode
- `bash K/scripts/study_reference.sh <out_dir> <url...>` downloads public videos (yt-dlp) for style analysis (sheets, beat/motion/luma metrics, loudness). Analyse for craft; do not redistribute or copy someone's brand.
README.md · README · 67 lines
- # make-explainer-animated-videos
- **Turn a voice-over and a script into a premium dark-mode animated explainer video** (1080×1920 Reels / Shorts / TikTok), with a cleaned voice, a properly mixed music bed, sound effects, camera moves and transitions. 100% free and open source. No footage, no AI video, no paid tools. Everything is rendered from code, so every frame is deterministic and tweakable.
- It is a [Claude Code](https://claude.com/claude-code) skill (an AI agent writes the storyboard and drives the scripts), but every step is also a plain script you can run and edit by hand.
- 
- *One storyboard, four looks (Mono Pulse, Blueprint, Amber Signal, Terminal Green).*
- ## What you get
- - **The look:** near-black stage, a ~4k-particle hero that morphs between shapes and words, thin 3D wireframe diagrams with glowing active parts and travelling packets, word-synced captions with underlined key terms, a progress rail, rolling counters, stamps, a recap list, a comment-CTA bubble.
- - **Voice first:** automatic cleanup of your recording (pauses capped, breaths and "ahh/umm" fillers removed, stretched vowels shortened, level-matched, speech chain), with an edit list that says what was removed and why.
- - **Sound:** a music bed picked for groove, warmth and energy (techno / house / tech-house, not pop), SFX by role (whip, impact, pop, tick...), and a mix **solved from the stems**: voice 10-14 dB over music + SFX, never closer than 8 dB, -14 LUFS.
- - **Energy:** camera snap/drift/punches, transitions only where the idea changes (each with its own sound), impact rings on reveals, a beat-locked glow, film grain, nothing static for more than 3 s.
- - **Honesty gates:** invented numbers are labelled "simplified example" unless sourced; a safe-zone check for platform UI; QA that measures loudness, balance, pacing and layout.
- - **Publish pack:** frame-0 headline, cover A/B, subtitles (.srt), a research-first posting kit with a linter, a LinkedIn post + 4:5 image, a publish checklist, and an upload script that **refuses to run without explicit approval flags**.
- - **Languages:** English, or Roman-script Hinglish (and any language whisper can't read, via acoustic-island timing).
- ## Requirements
- macOS or Linux (tested on macOS), **Node 22+**, **ffmpeg**, **Python 3.9+ with numpy**, **whisper.cpp** (`whisper-cli`) with an English model, a Chrome the renderer can use (downloaded automatically on first render). Optional: `librosa` (index your own music), `rclone` (upload), `yt-dlp` (study mode).
- ## Install
- ```bash
- # as a Claude Code skill
- git clone https://github.com/avijeetpandey/make-explainer-animated-videos ~/.claude/skills/make-explainer-animated-videos
- bash ~/.claude/skills/make-explainer-animated-videos/scripts/setup.sh # checks tools, downloads the whisper model, GSAP and fonts
- ```
- Standalone (no Claude): clone anywhere and run the scripts in `scripts/` (see `docs/USAGE.md`).
- ## Quick start
- 1. Record your voice (one sentence per breath, ~0.3 s pauses between sentences, quiet room) and write the script.
- 2. Tell Claude: `/make-explainer-animated-videos` + the voice file path + the script. Claude writes `scenes.json`, builds, checks, renders a draft and shows you the contact sheets.
- 3. Or by hand: copy `examples/sample/scenes.json` to `~/Documents/explainer-videos/my_video/`, add `voice.wav`, edit the text, then:
- ```bash
- S=~/.claude/skills/make-explainer-animated-videos/scripts
- python3 -I $S/build_explainer.py ~/Documents/explainer-videos/my_video
- cd ~/Documents/explainer-videos/my_video && npx --yes hyperframes@0.8.209 render -o renders/picture.mp4 -q delivery -f 30
- bash $S/finalize.sh renders/picture.mp4 renders/my_video_final.mp4 assets/audio/master.wav
- python3 -I $S/qa_explainer.py . renders/my_video_final.mp4
- ```
- Full walkthrough with every command: [docs/USAGE.md](docs/USAGE.md).
- ## Make it yours
- - **Handle, libraries, drive destination:** copy `config.example.json` to `config.local.json` and edit (git-ignored).
- - **Music and SFX:** bring tracks you may legally use (see [library/README.md](library/README.md)) and index them with `scripts/build_music_index.py`. No library = no music (it tells you); no SFX library = a small synthesised set.
- - **New look, shape, element, transition, language:** [docs/TWEAKING.md](docs/TWEAKING.md).
- - **How it works:** [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md). **Lessons learned:** [docs/LESSONS.md](docs/LESSONS.md). **Publishing method:** [references/PUBLISHING.md](references/PUBLISHING.md).
- ## Project layout
- ```
- SKILL.md the agent's instructions (workflow, rules, options)
- scripts/ build_explainer.py (main), engine.js + explainer.css (the renderer), clean_voice.py, audio_lib.py, fx_plan.py,
- mix_audio.py, qa_explainer.py, safe_zone_check.py, make_cover.py, make_srt.py, make_linkedin_image.py,
- seo_research.py, kit_lint.py, publish_check.py, upload_drive.sh, build_music_index.py, make_sfx.py, setup.sh ...
- assets/ worlds.json (looks), fonts/ (OFL), posting-kit-template.md
- references/ STAGE_LIBRARY.md (the storyboard format), STYLE_ANALYSIS.md, ENERGY_AND_SOUND.md, VOICE_CLEANUP.md, PUBLISHING.md, ROADMAP.md
- examples/sample/ a working storyboard
- tests/ python3 -m unittest discover tests
- ```
- ## Honest limits
- Wireframes are edges (no solid faces); one particle hero at a time; Canvas 2D, not WebGL; automatic filler detection only catches sustained sounds; Hinglish word timing inside fast stretches is +/-0.2-0.4 s (island starts are reliable; a hand-timed `words_override.json` always wins); the mix is judged by numbers, so listen once with headphones. The look was designed from an analysis of public explainer reels (see `references/STYLE_ANALYSIS.md`); **copy the format, never someone's brand or content.**
- ## Credits and licences
- Code: MIT (see LICENSE). Renders HTML with [HyperFrames](https://github.com/heygen-com/hyperframes) (Apache-2.0); timing with [whisper.cpp](https://github.com/ggml-org/whisper.cpp) (MIT); GSAP as a timeline clock (GreenSock standard licence, not OSI); fonts under SIL OFL. Details: [NOTICE.md](NOTICE.md). Music and SFX are not included.
- Built by [@avijeetpandey](https://github.com/avijeetpandey). Issues and pull requests welcome.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.