A video and explainer toolkit for AI agents
Start from an idea or a script and get a finished film with narration, subtitles, charts and a host. Every step is documented in plain language, and the core flow runs on your own computer with no account and no API key.
Voice, timing, visuals and assembly can all run on your own computer. The first example runs start to finish on free tools.
Voice, generated shots, host, editor: use the ones you already like. Follow the setup notes and they plug in.
In basic trials, three different agents read only the docs in the project and each made a film that passed the checks, including a local model running on a laptop and DeepSeek on its free tier.
Most AI editing tools trim footage you already shot and add subtitles. Script2Film starts from the idea: the film is built to match what you want to say.
The script is on the left; what it draws is on the right. When the narration says "beans", the first slice opens; at "biggest", that slice is highlighted.
Every image below is a real film frame or real tool output.
Each line is recorded several times, transcribed and compared with the script, and the closest take is kept and joined. Words that get misread go into a shared pronunciation table that every later film uses.
Line charts, pies, formulas, breakdowns, comparisons, lists, timelines, map routes, text highlighted passage by passage. In the script you write "appear when this word is spoken", and the program lines it up with the real timing of the voice. A quotation comes out of a blur word by word as it is read, a word is struck out and rewritten, a timeline slides from one stretch to the next, a clue board strings ideas together with red thread. 15 looks, and three of them can be tried side by side on this episode's own beats before anything is voiced.

The tool works out how long each shot must be, writes the prompts and lays them out as a run sheet. By default it submits to VideoExpress and collects the results; another platform, or generating by hand, works too if you put the files back under the expected names. Returned clips are measured and checked for bad first and last frames.

examples/brewing: a generated shot placed on the timeline, with the host circle and subtitles layered on as usual.The narration is cut into segments, sent for lip sync and placed back on the timeline. The host can sit in a corner circle, in a frame on the right, or fill the screen. You are shown a few look images first, and nothing is sent until you pick one.

The voice track is transcribed locally to get the time of every word, and both subtitles and visuals are laid out from that timing. A timed subtitle file is also written for later use.

Every exported film is checked on the spot, whatever kind it is: audio and video the same length, no black frames inside, loudness and peak, sound effects really in the mix, and made from the current files. The result stays next to the film with its SHA-256, and only that file can be uploaded. For explainers the first and last frame of every beat go into one sheet; the agent has to read it frame by frame before reporting done, and reports what it cannot make out instead of guessing.
Matters of taste are listed and never fail the film: music that does not step aside between lines, an opening that holds one picture for more than five seconds (the cover does not count), generated shots on so many lines that the high points no longer stand out, an episode too close to the three before it. The thresholds come from measuring published films, and where each came from is written in the code.
Background music is composed by code from the on-screen events, so sections start on the beat. Sound effects are lined up with the moment each element appears.
Before anything is generated or uploaded, the questions that need asking are answered one by one: rights, sources, whether anyone is depicted wrongly. Unanswered items stop the run.
Schedule, thumbnail, description and the altered-content label go up in one step. Whether to upload and when to publish is your call.
Measure the rhythm of a film you admire next to your own: shot lengths, how fast the opening changes picture, how much is speech, how quiet it gets between lines, plus a contact sheet with the narration under each frame. It measures; it does not judge for you.
The author uses it every week on two channels. Here is one film of each of four kinds; each opens on YouTube.
Channels: NotoriousLab (English) · Jacob Mei (Chinese)
Each stage reads the file the previous one left and writes its own. You can stop at any stage to look, edit and rerun.
Lists what is missing and how to install it.
Narration and the visual for each beat, in one file.
Voice every line and keep the most accurate take.
Measure when every word is spoken.
Code draws each beat to that timing.
Voice, visuals and subtitles become one film.
Measure, build the check sheet, read every frame.
Six optional stages attach alongside: generated shots, host, music and sound effects, overlays in an editor, checklists, and upload. The full steps and commands are in the flow document.
The trials ran on the author's MacBook Pro (M3 Max, 36 GB RAM). For all three agents and four models the film passed the output check, and its length fell within the range expected for the example (a 25-second English film).
| Agent | Model | Core flow (no account) | Generated shots + host |
|---|---|---|---|
| Claude Code | Claude Opus | Works | Works |
| Google Antigravity | Gemini 4 Argon, low | Works | Works |
| opencode + Ollama | Qwen 27b (MLX), local | Works, about 13 minutes | Works, about 35 minutes |
| opencode | DeepSeek V4.1 Flash (free tier) | Works, about 20 minutes for two films | Not tried |
Where each one got stuck, and how to get past it with another agent, is in the trial notes.
Verified so far on Apple Silicon Macs (tested on a MacBook Pro M3 Max). The first example is a four-line English film of about 25 seconds and needs no account.
Clone it with git, or download the ZIP and unpack it.
git clone https://github.com/notoriouslab/script2film.git
It only checks and installs nothing, and tells you what is missing and how to install it. A first install on a new machine takes about 30 minutes and about 7 GB of downloads; the steps are in the install guide.
python3 tools/doctor/doctor.py
Then tell it: