Building an
Autonomous Startup
The thing about Claude Code is that it isn't really a magic model. It's still using Sonnet or Opus under the hood, which are great, but they're not going to do the things that Claude Code can do. Claude Code's magic is like a magician's trick—it looks incredible, but it's surprisingly simple once you see how it's done.
I always told people programming is just if statements and loops, and if you want to program, you don't need to learn anything else. It's an exaggeration, but these are the fundamental concepts that run nearly everything digital. Claude Code is a reinforcement of this idea. The command line tool calls Anthropic's models in a loop with a lot of clever prompting to keep the model running in a constructive way based on limited user input.

Spoilers: this is what it built. You can visit the startup it built at claude.dwyer.co.za or see the GitHub project at github.com/sixhobbits/claude-experiments.
The first thing I thought of doing was extending that loop—maybe infinitely? How far could Claude Code go if it was given a few resources, like a root VPS, and some minimal instructions to never terminate and to just go forever?
I told it to build a startup. I had some issues getting Claude to make sure that Claude never terminates, so instead I had Claude write a simple while loop bash script that keeps calling claude -p
(more on this later, but -p
is a really powerful way to use Claude).

It wrote its own prompt (link). I created a root VPS and told it to go.
Aside: I hit a small snag where Anthropic decides that running Claude as root with --dangerously-skip-permissions / yolo-mode is not allowed. You can get past this dumb nanny-state stuff by running:
export IS_SANDBOX=1 && claude --dangerously-skip-permissions
It evaluated a bunch of startup ideas, rated them, and got to work.
The idea it came up with (server monitoring) doesn't make any sense at all and it never realised it. It's a web app, so the only server it can monitor is the one it's running on, but from the copy it seems to think it's a SaaS tool you can sign up to and monitor your own servers. You can't.
BUT this is still seriously impressive stuff. It configured a fully working full stack web application, including Nginx, certificates, etc etc. It's doing real (if misguided) development work, with nearly no input from me at all.
Aside: Most people I know would criticize this in the same way that AI has always been criticized. "It's not real bro, it's just pattern matching. It's seen stuff like that before. It's not even working properly. An intern could do that with a bit of time."
The thing about these criticisms is I've been hearing them since I got into character-based neural networks in 2015. The criticisms never change, it's just the line that moves.
Or
Whatever, I don't want to get into the debate too much here, but a) I am fucking impressed. b) I would never have predicted an artificial system that could do this 10 years ago or even 6 months ago, and c) anyone else who claims otherwise is likely lying or has ulterior motives. Have a nice day.
For about a day it kept building features even though the log-in page errored out and nothing worked.
I gave it a few nudges via my HUMAN_INPUT.md file and it got a basic demo up and running, a nice looking landing page, and a few other nice touches like self-referencing that the project was built by autonomous AI.
Hitting a Snag
Most of the time I could interact with it just by:
- Seeing the changes it made to the production website
- Seeing the outputs it added to GitHub in the various note files and what human help it asked for
- Adding stuff to HUMAN_INPUT.md
I never needed to SSH into the VPS until it stopped working. After 6 hours of no commits I had to login to check what was happening:
[Fri 25 Jul 2025 02:29:41 AM UTC] Starting Claude process... API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy (https://www.anthropic.com/legal/aup). Please double press esc to edit your last message or start a new session for Claude Code to assist with a different task. [Fri 25 Jul 2025 02:29:47 AM UTC] Claude process exited with status: 1 Waiting 3 hours before restart..
Uh oh. We're getting blocked again and I've heard Anthropic has a reputation for shutting down even paid accounts with very few or no warnings.
I read the User policy and saw that my recent inputs telling it to go ahead and market the startup to get users had probably tripped some big brother switch. The user policy (which obviously I read before when I agreed to it) states that automatically published content needs a human-in-the-loop and Claude was trying to promote the startup on Hackernews without my sign off.

I did what it asked (posted its stuff as-is to hackernews and reddit, luckily I didn't get banned from either but I did get ignored by both), and then watched it talk shit about user acquisition for a bit. It got lost trying to monetize through a free trial and social proof stuff, which was directionally correct even if completely non-sensical in context and then decided to turn it off so I could save my (still limited, even under the max plan) usage for some more useful stuff.
The project has 100 commits, so if you want to see exactly what it did, you can take a look at each of those.
(More in a later section, but it's also my text editor. Look at me writing this article in Claude Code.)
