The video above was produced with OpenMontage during the walkthrough itself. This article is its practical companion: you can watch the complete process, copy the production brief, and use the written steps while building your own version. In the demonstration, I create an approximately 80-second documentary called What If the Internet Disappeared for 24 Hours? The workflow researches the topic, proposes a story, writes narration, plans scenes, collects public assets, generates a local voice-over, composes the edit, and exports an MP4 for review. This is also one building block in my larger experiment: a personal YouTube agent that can coordinate video production and, after explicit approval, prepare work for the Tomorrow Without YouTube channel. The tutorial demonstrates production and export; it does not claim that unattended YouTube publishing is complete. What OpenMontage does OpenMontage is an open-source, agentic video-production system. Instead of generating one isolated clip, it gives an AI coding agent a structured production pipeline. The workflow can coordinate: live research and source collection proposal and script development scene planning and editing direction public or generated photographs and footage local or API-based narration music and ambient sound Remotion or HyperFrames composition review checkpoints and final rendering The important distinction is that the coding agent acts as the orchestrator. OpenMontage supplies the production conventions, tools, pipeline stages, and quality gates. What you need Before starting, install: Python 3.10 or newer Node.js 18 or newer FFmpeg an AI coding assistant that can inspect files and execute commands I used Codex as the coding agent and ran the demonstration on an M4 Pro Mac with 24 GB of unified memory. That is context from my test, not a stated minimum requirement. OpenMontage can use optional providers such as ElevenLabs for premium voice-over. In this walkthrough, Piper handles text-to-speech locally, so narration does not require a paid voice API. Step 1: Clone and set up OpenMontage Clone the official repository and enter the project: My first SSH clone attempt failed during the recording, so I switched to the HTTPS clone URL above. If make is unavailable, follow the manual setup instructions in the repository README. The setup installs the Python dependencies, the composition runtime, Piper for local narration, and the supporting production tools. OpenMontage can select Remotion or HyperFrames based on the proposed video and lock that choice for the production. You can run the setup yourself, or ask your coding agent to inspect the repository instructions and perform the documented setup. Either way, review what it plans to install before approving commands. Step 2: Write a production brief, not a vague prompt The result depends heavily on the production brief. Define the audience, narrative structure, visual language, tone, evidence standard, duration, and output format. Here is the reusable prompt behind the demonstration: For a Short or Reel, change the duration and explicitly request a vertical format. In this demonstration, I requested landscape output for YouTube. Step 3: Let the agent research before it writes After receiving the brief, the agent researches the subject and stores the source material with the project. In this run, the research supported a story about cascading disruption rather than immediate societal collapse. That distinction improved the script. Online communication, cloud access, digital payments, logistics, and other connected systems would be affected differently over time. The production therefore follows a chronological structure instead of making one unsupported catastrophic claim. Do not treat collected media as automatically safe to use. Review source provenance, usage rights, factual relevance, and whether each asset actually supports the narrated claim. Step 4: Review the production in Backlot Backlot is the local production board where you can see the project state and approve work. Open a production with: The demonstration pauses at the proposal gate before continuing. From Backlot, you can inspect the planned format, duration, research direction, script, scenes, assets, edit, composition, and publishing stage. The generated hook in this run begins with ordinary failures: a message that will not send and a checkout that stalls. That is more relatable than starting with an abstract explanation of network infrastructure. You can approve stages one at a time or ask the agent to revise a specific output. If you already have a complete script, provide it and ask the workflow to restructure it for production rather than generating a new one. Step 5: Approve each production stage The practical flow is: Research — inspect the sources and narrative findings. Proposal — confirm audience, format, duration, tone, and runtime. Script — read the narration and correct weak or unsupported claims. Scene plan — confirm that every scene communicates a useful idea. Assets — review footage, photographs, prompts, rights, and relevance. Edit — inspect pacing, transitions, overlays, music, and sound direction. Compose — preview the assembled video and test the result. Publish or export — approve the final delivery only after review. These gates are not wasted friction. They let the agent automate repetitive production work without hiding important creative and factual decisions. Step 6: Review the rendered MP4 At the final composition stage, OpenMontage exports an MP4 for review. Check more than whether the file opens: listen for pronunciation and awkward narration confirm the voice matches the intended serious, curious tone verify that every factual claim has appropriate supporting context look for irrelevant, repetitive, or low-quality visuals inspect text legibility and safe placement check music and ambient sound against the narration confirm the video duration, resolution, and aspect ratio The first render in the demonstration was usable, and the local Piper narration was clear. I still made one visible revision: adding the watermark Made using OpenMontage by Praveen Kumar Rana. You can use the same revision loop for a logo, title treatment, or other brand element. What the demonstration proves This walkthrough shows that one well-scoped brief can drive a complete, reviewable production from research to an exported documentary video. It also shows why the approval workflow matters: the result is created through several inspectable decisions, not one opaque generation request. It does not prove that every topic will produce an equally strong video, that every public asset is automatically licensed for your use, or that creative review is no longer necessary. It also stops at a reviewed export rather than demonstrating unattended publication to YouTube. Common failure points Clone or setup fails: use the HTTPS repository URL and follow the current README for your operating system. The prompt is too vague: specify audience, duration, aspect ratio, evidence standard, story structure, visual direction, and tone. Research becomes sensational: explicitly request credible sources and prohibit unsupported catastrophic claims. Assets feel generic: ask for realistic full-screen documentary material and review every selection. Voice-over needs improvement: revise narration direction or configure another supported voice provider. The first render is not final: treat it as a review artifact and request targeted changes instead of restarting blindly. Where this experiment goes next This documentary workflow is the foundation for the broader autonomous YouTube-agent experiment. The next capabilities I am exploring include advertisement workflows, Shorts and Reels, reusable custom skills, specialized agents, stronger automated quality checks, and governed YouTube publishing. The goal is not automation for its own sake. It is a repeatable production system that preserves sources, creative decisions, approvals, assets, and final deliverables—while making it clear where human judgment is still required. You can follow the resulting experiments on Tomorrow Without and use the OpenMontage repository to reproduce the workflow.