Scaffolding for one
Series: Scaffolding for one (0/6). Intro post.
I am one person. Voice Legacy is in production. Real people pay for it, real family stories live in its database, and every day the system does work that I used to do by hand.
The question I get most often from other PMs trying to build with AI is some version of “how are you shipping this alone?” My honest answer is that I spent the first few months of this project building the boring operational layer around the product, before I built most of the product itself.
This is a series about that boring layer: the scripts, the skills, the cron jobs, the eval framework, the backups, the dashboard, and the agents. None of it is the Voice Legacy product itself, and the product would not survive the week I am having right now without all of this running in the background.
I am writing it for two reasons. The first is that every solo founder I know is reinventing these pieces from scratch, badly, under time pressure, in the middle of trying to ship features. The second is that this layer is the single biggest lever I have found for keeping one person productive while running a real AI product. If you build it early and cheaply, it pays you back every single day.
The series covers six topics.
Marketing. I walk through how I write once and distribute everywhere, including the Claude Code skill that runs the whole pipeline and the two-repo pattern I use to organize my workflow tools.
Quality. I break down the eval framework with two test personas and LLM-as-judge scoring, the Playwright suite that drives the actual UI, and the CI pipeline that runs both on every pull request.
Data. I show the GitHub Actions that back up the database and the media bucket nightly, with lifecycle policies so I am not paying full price for data I will never read.
Claudia Control Plane. I give you a tour of the Next.js dashboard I built as my chief of staff, the Postgres job queue behind it, and the Mac mini that runs the actual work.
The Dreamer. I describe a weirder agent in the control plane that does garbage collection on its own memory, analyzes its own job history, and writes a first-person weekly reflection on what it has been doing.
What I would tell a PM trying to do this. I close by listing the things I am still missing, the mistakes I have already made, and the one piece of advice I would have wanted three months ago.
Every piece of this is in one of two repos. The file paths are real, the numbers are real, and where something is half built or aspirational I will say so. I am not trying to sell you on a methodology. I am showing you what one person can actually run.
Post one goes up tomorrow: how I write once and publish everywhere.