Point it at a blank prompt and ask for a video about a topic, and you will get something competent and generic. That is the trap. The value is in the repurposing pipelines, the ones that start from footage that already exists, because that is where the machine does the tedious work and your taste still decides the output.
So the framing I would give you: this is a rough-cut assistant, not a creative director. Feed it long footage you already made, let it do the passes that make your eyes glaze over, and keep the final cut on your desk.
brew install ffmpeg on a Mac, sudo apt install ffmpeg on Linux, or grab it from ffmpeg.org.git clone https://github.com/calesthio/OpenMontage.git. This downloads the whole thing to your machine.cd OpenMontage. Everything from here happens inside that folder, which is also the folder you will open in Claude Code.make setup. This builds a Python environment, installs the dependencies, installs the Remotion rendering engine, installs the free Piper text-to-speech voice, and creates your .env file. It takes a few minutes and prints a lot of text, which is normal.npm install fails on Windows with an ERR_INVALID_ARG_TYPE error, the documented fix is to run npx --yes npm install instead.The repo ships with its own instruction files, and the difference between a good first session and a confusing one is whether your assistant read them. Make that the first thing you ask for.
Read AGENT_GUIDE.md and PROJECT_CONTEXT.md in this project before you do anything else, then tell me what I'm working with in plain English, assuming I'm a creator and not an engineer.
Specifically: list the pipelines available to me, say which of them work right now with no API keys configured, and name which free tools are doing the work in that no-key path. Then tell me the single pipeline you'd recommend I try first if my goal is [cutting my long videos into short clips / editing my own talking-head footage / making an explainer from scratch], and why that one.
Do not start a production yet. I want the lay of the land first.This is the pipeline to test first, because it is the one where the output is judged against footage you already trust. Have a long video ready: a webinar, a podcast episode, a long-form YouTube video, anything over ten minutes.
Use the clip factory pipeline on this source video: [file path or link]. My goal is vertical short-form clips for Instagram and TikTok, each between 30 and 60 seconds.
Before you cut anything, transcribe it and give me a ranked shortlist of the ten strongest moments, with a timestamp, a one-line reason each is strong, and a suggested hook line for each. I want moments that are a complete thought on their own, not a sentence that only makes sense if you heard the previous five minutes. Wait for me to pick which ones to build.
Then, for the ones I choose, cut them, add word-level captions, and render them vertically. Show me the ranking logic you used so I can tell you what you got wrong about my audience.The system's knowledge lives in readable markdown files, over 700 of them, in a skills/ folder. That is the customisation surface. You are not editing code, you are writing down your editing preferences in plain English so the agent applies them every run instead of you re-explaining them.
I want you to capture my editing style as a reusable style file in this project, so every future production follows it without me repeating myself.
Here's how I edit: captions are [style, position, font feel], cuts are [tight with no dead air / breathing room between beats], my hook is always in the first [X] seconds and it is [a question / a claim / a result], I never use [stock-looking B-roll / trending audio / zoom transitions], and my videos end on [your sign-off]. My audience is [who they are] and the tone is [warm and practical / fast and punchy].
Read how the existing style and skill files in this repo are structured, write mine in the same format and the correct location, and then tell me exactly which pipelines will pick it up and how I'd change it later.This is the part I want to be exact about, because "free AI video" usually means "animate some still images and hope". Here, with zero API keys configured, the setup gives you a genuinely working stack.
The paid layer is a long list of optional generation providers you add as API keys in the .env file, and every one of them bills you directly. None of them are free tiers dressed up, they are pay-as-you-go accounts you open yourself. The repo publishes real costs for its own showcase videos, including a 60 second animated short it produced for $1.33 and a 50 second vertical film for about $4, which is a useful reality check on what generated footage actually costs. The full provider list with pricing is in the repo's providers document.
One feature genuinely worth knowing about: it can take a reference video, a YouTube video, a Short, a Reel or a local clip, analyse its transcript, pacing, scenes and style, and come back with differentiated concepts, an honest tool path and a cost estimate before it produces anything. For anyone who finds a blank prompt paralysing, that is a much better on-ramp than inventing the perfect brief from nothing. The repo's prompt gallery has tested prompts with expected costs if you want more starting points.
OpenMontage on GitHub, the repo itself (prerequisites, install commands, pipelines, zero-key tool table)
The AGPLv3 licence it ships under
AGENT_GUIDE.md, the instruction file your assistant should read first
The provider list with pricing and free tiers
Claude Code docs, for installing the assistant that runs it