Solo Founder + AI: I Finally Shipped Three Things I'd Been Putting Off for Ages
Over the past 7 days I logged 400+ AI collaboration sessions and 90+ git commits. The numbers don't really matter. What I want to talk about are the three things that actually solved real problems, and the mistakes I made along the way.
1. Mystery shoppers for restaurant owners
Lots of owners want to know what service at their outlets is really like. The old way was to ask a friend to go eat there, get a quick verbal rundown afterward, and quietly pay them back. It wasn't a proper process, and nothing got recorded.
This week I added a mystery shopper flow to the product. The merchant generates an invite link and sends it over. The shopper opens it, picks an outlet, works through a questionnaire item by item, and uploads photos from the visit plus the order receipt. Once an admin approves the submission, the area manager reimburses the meal to the bank account or PayNow number the shopper provided.
My process: I had the AI draft a complete plan first, then went through it point by point. After that I had it split the work into a few commits by feature, then merged and deployed. While I was at it, I added backups to the scheduled jobs too. Right after launch I spotted a problem: each link only lets you pick fixed dates and time slots, which is a lot less flexible than a proper booking flow. That's the next thing to fix.
2. Deleting production data: my three rules for the AI
The system still had a batch of test onboarding records and accounts. Testing was over, so they had to go. The scariest part of deleting production data is deleting the wrong thing.
So I set a process: plan first, then back up, then do a dry run, and only execute once every single record on the delete list checks out. During the dry run I noticed one test staff account had been left off the list. I added it, and only then ran the real deletion. Afterward I wrote this one-off exception up in the docs.
It made one thing clear to me: the AI executes fast, but I'm the one who decides which steps need a human to stop and confirm.
3. Automated marketing is up and running, and rate limits taught me a lesson
Content generation, images, translation, and multi-platform publishing are now basically one end-to-end pipeline. I also switched image generation to an open-source model running on my own machine.
Plenty went wrong, too. One batch of posts got zero views, most likely because they were throttled, so I had to review them and repost. The browser extension's permissions were misconfigured, and for a while I couldn't even get into the WeChat Official Account dashboard. On Xiaohongshu I split things into a product account and a personal account, and the personal one now posts automatically every day on a scheduled job. On the international side, I've submitted an application for LinkedIn's Ads API. I looked at TikTok and decided to hold off on applying for now.
Also: a side project just for fun
Outside of work, I'm using several AI sub-sessions in parallel to build a small open-world game set in Singapore. This week I added street racing and shooting competitions, and every module has automated simulation tests to check the game balance numbers. I'm using it to practice splitting a big task across multiple agents working at once.
Wrapping up
The biggest takeaway this week isn't how much code got written. It's that I've figured out how to split the work with AI: it brings the speed, and I set the rules, make the judgment calls, and catch what slips through.
What's the most genuinely useful thing you've done with AI at work? Drop a comment, I'd love to hear it. And if you run a restaurant in Singapore and the mystery shopper feature sounds interesting, feel free to DM me.