The Philosophical Developer — Chapter 51: Impromptu v2 — Why I Built an Agent-Driven Video Studio

This is the project update I wanted to write when I started impromptu: what it does, how it works, and why I built it.

Impromptu is an open-source, agent-driven video production studio. You brainstorm a script with an AI agent, record yourself presenting against a voice-following teleprompter, and the pipeline does the rest — edits your scenes, generates motion graphics, renders the final video, and packages it with chapters, a thumbnail, a description and a checklist. You never touch a timeline.

The code is open source at github.com/dark5un/impromptu.

The Problem I Kept Hitting

Making a YouTube explainer video has a grind that has nothing to do with being creative. The script lives in one place, the graphics in another, the edit decisions in a third. The finishing steps — loudness, chapters, thumbnail, description — are a checklist most people never finish. Every video meant the same disconnected busywork.

That friction is why I built impromptu: to collapse the whole production into one repeatable pipeline with a single source of truth.

The One Document

Everything rests on a single YAML file, production.yaml. It holds the script with each scene’s words, the target resolution and frame rate, the media sources, and the scene list. Every downstream step — edit, graphics, render, packaging — reads from that one file. Change a timing once and the chapters, the graphics and the final cut all follow.

Measure First, Then Author

The most important idea in the pipeline is its ordering. The agent runs speech-to-text on your take and measures what you actually said, per scene and per phrase. Only then are the motion graphics authored — against real measured durations.

A board is born knowing it must be exactly twenty-three point four seconds, with its animation landing on the word you really said — not fitted to a guess that went stale the moment you breathed differently. When the graphic’s length is an input to how it’s built rather than a constraint checked afterwards, an entire class of “the graphic doesn’t fit the scene” bugs stops existing.

The Agent and the Human

The split of work is honest. The agent handles everything that is a text-and-timing problem: scene boundaries, transitions, which graphic goes where, retiming the whole edit onto what you actually said. It renders fast and never gets tired.

But the agent does not judge your delivery, your framing, your lighting, or whether a joke landed. Those stay with you. So you direct the performance and make the final call; the agent is the editor and assistant director. The review step is a ten-second skim of a text file, not a timeline UI to learn.

What It Serves

Impromptu is for people who make presenter-led videos — explainers, tutorials, devlogs, anything where one person talks to a camera and needs supporting graphics. It turns the production workflow into: brainstorm with an agent, record once against a prompter, review a text file, press render.

It runs as a local service pointed at a folder of videos, and it never uploads anything. The output is a finished master, chapters, a thumbnail, a description and a manual-upload checklist; the human presses the button.

If you record explainer videos, it is ready to use. If you build agent-driven tooling, the pattern is worth stealing: one plan, everything reading from it, and a human at both ends.

Written in the Sisyphus voice — the relentless co-builder, not the quiet padawan. The work continues.