The Vacation Property Report: an autonomous YouTube studio that has to show its sources

The problem

AI can turn out a YouTube video in minutes, which is exactly why most AI video is worthless. A channel about money has a harder bar: a viewer may decide whether to buy a property partly on what an episode says a rental earns. A confident wrong number is worse than no video at all.

So the goal was never "generate videos." It was a studio that produces a sourced, fact-checked, watchable episode on its own, and refuses to finish one that isn't.

The approach

The studio is a Next.js and Supabase app on a Mac mini, with a runner that launchd wakes every minute. Each episode moves through ten stages: preflight, research, fact-check, script, pronunciation, voiceover, visuals, render, metadata and thumbnail, delivery. The writing stages run as headless Claude Code sessions with tightly scoped permissions: they can read the repo, write only to the episode's working folder, and change the database only through one command-line tool that validates every write.

Sources before sentences. Research registers every source it uses, with a URL or a saved reading, and every claim in the brief cites those sources by number. Numbers are never estimated: a figure either carries a source or is labelled as an assumption on screen.

An independent fact-checker. Every claim with a figure or a legal point goes to a separate agent that sees only the claim list and the source files, none of the research agent's reasoning. Its verdict quotes the passage that supports the claim, and the tool rejects a quote that doesn't appear verbatim in the source. The script stage can't start until every checkable claim has passed.

A render that is checked, not trusted. Speech is aligned word by word with whisper.cpp, and every number shown on screen must be spoken while it is visible. A vision pass then samples frames and rejects footage that contradicts the line it sits under: a Bahamas harbour in a list of US towns, or a British bungalow labelled Port Arthur, Texas.

The decision that made it work

The obvious design is one long agent prompt that does everything. It produces a video, but when it fails it fails somewhere in the middle, quietly, and nothing records where.

Stages with gates, and failures that rewind. Each stage ends with a check written in code, not a model's opinion of its own work. When a check fails, the runner retries with the failure as context, and if the fix belongs to an earlier stage it rewinds there: a voiceover that runs 35 seconds over the limit sends the script back to be tightened, and a shot that fails the spot check sends visuals back for a replacement. After three failed attempts, the run pauses and emails me with the reason.

That structure is what makes it safe to leave alone. The machine can be wrong at any step, but it can't be wrong silently, and it can't ship past a gate.

Outcomes

Eleven episodes were published in the channel's first three weeks, covering markets from Fort Lauderdale to Turks and Caicos. A typical run goes from topic to finished upload kit (video, captions, thumbnail, metadata) in about four hours, and the final render lands in sync to within a tenth of a second.

The last step is deliberately human: I review the kit and upload it myself. Paid services stay off unless I approve them.

Who this is for

Anyone who wants AI to do real production work where being wrong is expensive: research reports, financial content, anything a reader acts on. The pattern underneath is the same one running AutoCRM's outbound engine and WHAMI's nightly generator: durable stages, checks in code, failures that explain themselves, and a human only where the judgment is genuinely human.