Skip to content
Projects

anneheartrecord/charles-video-maker

View on GitHub

anneheartrecord/charles-video-maker has 90 GitHub stars and 11 forks. GitHot measured +7 in 7 days. Last checked 2026-11-18 11:24 UTC. Last push 2026-11-04.

How these counts are measured

Measured star growth

in 24h
Update delayed
in 7 days
+7
in 15 days
Collecting history
in 1 month
Collecting history

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

Agent skill: script to narrated vertical video — code animation or real footage underneath, voiceover, subtitles, fact-check, Anatomy machines, optional Lemo-Opuscar styles

Creative & Media

README_EN.md

Original

charles-video-maker

charles-video-maker

A script goes in; a finished video with voice, subtitles, cover and publish copy comes out. The picture is drawn by code, frame by frame.

English | 中文

License: MIT Python Node FFmpeg Agent Skill

A zero-follower account, 13 days, 370K views

On 2026-10-20 a new Douyin account was started with this skill's Douyin mode. No paid promotion. A person picks the topic and does the final review; everything from script to publishing is done by the pipeline.

Total views Best single video Most shares on one video Ad spend
370K+ 231K 7,053 0

Showcase

The least happy year Six wins Franklin effect
At what age is your life least happy? Feel you're worse than everyone? Maybe you just haven't won 6 times Want someone to like you? How the Franklin effect works
78 s · 231K views 66 s · 99K views 71 s · 13K views
7,053 shares · 2,537 comments 1,441 saves · 102 new followers 373 likes · 178 saves
▶ Watch on Douyin ▶ Watch on Douyin ▶ Watch on Douyin

Numbers as of 2026-11-02. The same pipeline has also produced videos with a few dozen views: reach is mostly decided by the topic and the first 3 seconds; the tool keeps the finished quality at a steady level.

Start in 30 seconds

npx skills add anneheartrecord/charles-video-maker

Then tell your agent what you want:

Turn this article into a 60-second vertical video with voiceover and subtitles

What it does

Capability In one line Good for
Code animation The model writes a page that draws frame t; frames are captured and composited into a video explainers, data stories, product intros
Footage underneath Recordings, source-video quotes, royalty-free b-roll and generated stills sit under a transparent code layer product demos, game intros, commentary
Machines that explain A technical idea built as a working isometric machine that does each step as the narration reaches it deep technical explainers
Douyin knowledge-video mode Start from a video that is taking off, then topic, style, script, score, render, publish and stats run end to end growing an account, batches of short explainers
Style library (optional) The 53 styles of Lemo-Opuscar plus sampled-instrument scores films that need a specific look

Every video passes quality gates: every number must be in the fact-check table, every scene is checked on a contact sheet, and reading time, frame count, loudness and true peak are verified.

How a video is made

script ──► fact-check table ──► voice (one take, sped up) ──► word timestamps ──► scene cuts, cues, subtitles
                                                                                     │
          ┌──────────────────────────────────────────────────────────────────────────┘
          ▼
   code picture render(t) ──► frame capture (parallel) ──┐
   footage layer footage.json ──► cut, grade, fade ──────┼──► composite ──► out/<topic>.mp4 + cover + publish.txt
   score + effects + voice ──► -14 LUFS ─────────────────┘
  • Voiceover pipeline: the narration is synthesized in one take and Whisper word timestamps are aligned back to the script, so scene changes, actions, sound effects and subtitles land on the spoken words, and recognition mistakes never reach the subtitles.
  • Footage layer: record.py records a web page, source.py downloads and transcribes a source video and finds where a sentence is said, stock.py searches royalty-free b-roll, and compose.py composites footage, the transparent code layer and the voice.
  • Delivery: the video and cover are named after the title; publish.txt holds only the title, the description and 5 tags.

Douyin knowledge-video mode

All three showcase videos came from this mode. It does not start from an idea; it starts from a video that is taking off. The style is matched one to one, the content is never copied, and picture, texture and timing are pushed past the original. One request authorizes the whole run; it stops to ask only when no primary source can be found or the same tool fails twice in a row.

D-1 topic radar ──► D1 distill style ──► D2 sources & script ──► D3 score first, then picture ──► D4 build & self-check ──► D5 deliver
 douyin_radar        style library        gemini_copy             voiceover-pipeline                still_check               cover, publish.txt
                                                                                                                                │
                          findings written back to the ops handbook ◄── douyin_stats / douyin_comments ◄── douyin_publish ◄───┘
Step What it does
Topic radar douyin_radar.py ranks the videos rising under a hashtag and produces 10 to 20 topics
The first 3 seconds douyin-opening.md takes apart the openings of top videos, with checklists for picture, first line and sound
Script gemini_copy.py hands narration and title to a local Gemini; the skill checks facts and structure
Still check still_check.py finds stretches where the picture stops for more than 2 seconds, and subtitle changes the picture doesn't follow
Publishing and stats douyin_publish.py schedules the video, douyin_stats.py pulls 2-second bounce, 5-second completion, likes, saves and shares, douyin_comments.py saves comments
Read-back check say_check.py compares the script with the transcript and lists numbers and words the TTS swallowed or misread
Opening mix mix.py --open-boost --open-lift makes the first 3 seconds about 3 dB louder while the music stays clear of the first line
Styles and templates styles/ holds five distilled looks, templates/horizontal-explainer/ a landscape skeleton, templates/douyin-cover/ 3:4 and 4:3 covers
Batches douyin-batch.md: one sub-agent per video; the main agent assigns, reviews and schedules one at a time
Series douyin-series.md records the series approach and the current one-video-at-a-time rule

You set up two things yourself: a working directory for source videos, the style library and projects (env DOUYIN_WORKDIR, default ~/Desktop/video/vibe-video/), and an ops handbook you maintain for topic rules and findings (env DOUYIN_HANDBOOK). The Douyin login state lives where DOUYIN_PROFILE and DOUYIN_COOKIES point, by default under ~/.config/douyin/, and never leaves your machine. Without the handbook, videos still come out; the read-the-handbook and write-back steps are skipped.

More results

Dial indicator (Anatomy's smallest example) Raptor engine (with a WebGL plume)
Dial indicator video Raptor engine video
The carriage slides when the narration says so; the probe drops onto each gauge block as it is named Hover takes it apart into 11 assemblies; on ignition the WebGL plume fires in the same isometric camera

Four choices at the start

In the default mode the skill asks four things at once, each with a default; the Douyin mode skips this step.

Choice Default Alternatives
Picture code picture footage underneath, or one style from the library
Format vertical 1080×1920, 30 fps, cover as frame 0 horizontal 1920×1080, 24 fps, cinematic pacing
Voice Gemini TTS at 1.25x edge-tts, Kokoro (offline), none
Workflow three plans with previews, build after approval ask once, write a TREATMENT.md and build

More prompts

Make an intro video for this web game from real gameplay; the link is https://...
Here is the source video and my script; use the source and b-roll, not just illustrations
Make a vertical video on CDN cache hits and misses, explained by a working machine
A 45-second watercolor film about tea, from picking to brewing
Find 10 Douyin topics rising under #Vibe知识大赏 and make one from the best

Install and requirements

npx skills add anneheartrecord/charles-video-maker
# or copy it manually
git clone https://github.com/anneheartrecord/charles-video-maker.git
cp -r charles-video-maker/skills/charles-video-maker ~/.claude/skills/
Dependency Required / optional Notes
Python 3.9+ required pip install numpy scipy playwright, then playwright install chromium
FFmpeg required compositing and audio
Node.js 20.10+ required Anatomy builds and checks, built-in modules only
Voice one required GEMINI_API_KEY (Gemini TTS), or the free edge-tts
Whisper required mlx-whisper on Apple Silicon, faster-whisper elsewhere
Footage layer optional yt-dlp, PEXELS_API_KEY or PIXABAY_API_KEY
Douyin mode optional a Playwright browser profile logged in to the Douyin creator center (QR login on first run), Antigravity CLI agy
Style library optional sh skills/charles-video-maker/scripts/lemo/setup.sh deps, about 350 MB, samples about 1.3 GB more

Keys are read from environment variables only and never written to any file.

Layout

skills/charles-video-maker/
  SKILL.md        routing (video / figure / machine explainer / Douyin) + four choices + the video workflow
  references/     plans, brief, craft rules, cover and opening, voiceover, fact-check, machines, footage, styles, Douyin mode
  scripts/        rendering, fact-check, QA sheet, audio tools, Douyin radar / publish / stats / comments, Gemini script, still check
  scripts/voice/  voice, alignment, read-back check, timeline and subtitles, mix, sampled scores
  scripts/footage/recording, source quotes, b-roll, generated stills, compositing
  templates/      vertical and landscape explainer templates, Douyin covers
  styles/         Douyin style library (five looks)
  bridge/         virtual clock, video overlay, page wrapper
  anatomy/        Anatomy, vendored unmodified
examples/         images for this README

Credits and license

  • The video workflow, references and render scripts started from xilo-opus-video by Kianzzz.
  • anatomy/ is Anatomy by Ryan (@wheresryan22), vendored unmodified.
  • The Douyin mode's account-building and remake playbook is adapted from chapters 5 and 6 of WEILANX's Opus 5.5 Douyin account-building tutorial.
  • The style library, sampled-instrument scoring, Kokoro voice and reading-time check come from Lemo-Opuscar by LemoLab, downloaded on demand and not copied into this repository.
  • Sources, versions and changes are listed in NOTICE.md; plans are in ROADMAP.md.

MIT, with all original copyright notices kept. See LICENSE.