Build log/2026-10-28·4 min read

A Week of Letting AI Do the Work: What I Most Want to Share Is How to Stop It Making Things Up

#buildlog#building-in-public#Indie Hacking#AI Collaboration#Side Project Diary#ClaudeCode#One-Person Company#AI Coding

I ran the numbers this week: over a hundred tasks done with AI and 50+ commits. The count isn't really worth bragging about. What I actually want to talk about is this: the more work AI does, the more you need guardrails that can stop it.

Here are the main threads from this week, including what moved forward and what bit me.

1. My browser city got an airport

I'm building a 3D city game that runs in the browser. This week was all about the airport and flying:

  • I put together the control tower, hangars, lights and signage from free open-source models. All 21 models come to about 1.2MB once compressed and bundled;
  • You can rent a plane and buy insurance, or buy your own plane, with parking stands, maintenance and upgrades;
  • There are four kinds of flying jobs: sightseeing tours, aerial photography, airdrops and business charters. You get paid on landing based on how well you flew;
  • I also added airspace violations and a side storyline about aviation.

The bug I'll remember: at night, the green lights at the runway threshold were too bright. Once bloom amplified them, the whole end of the runway looked like it was sitting in green fog. I fixed it by turning the lights down a notch and shrinking the light housings.

The practice most worth reusing: for final acceptance, I had the AI fly a full approach and touchdown on autopilot at each of the three graphics settings, as a regression test. It also re-ran the downtown performance checks to make sure nothing got slower. I used to fly these tests by hand. Now a machine does it.

Also, someone on X posted city scenes with incredibly realistic detail. Before deciding what to borrow, I had the AI figure out how they actually did it. The answer was to layer a photo-scanned detail texture onto nearby building facades, and only switch it on at the high graphics setting so it doesn't slow down low-end devices.

2. MakanCloud: hardware bundles are live on the website

When I talk to small shop owners, the first question usually isn't about the software. They want to know which machines to buy and what it costs in total.

This week I had the AI research what other vendors offer for hardware and how they price it. Then I put three one-time-purchase hardware bundles on the homepage and pricing page: hawker stall, single outlet and multi-station. The terms are spelled out plainly: one-year warranty, no extra GST, and delivery in about a month once the order is confirmed.

3. AI writes, AI reviews

I've turned the MakanCloud blog into a semi-automated pipeline. One AI writes drafts from a topic list. A second AI acts as a compliance reviewer and checks each draft against 13 red lines, one by one. For example:

  • No made-up customer data. Any numbers in a case study have to match the published case studies;
  • Hypothetical numbers in examples must be clearly labeled as hypothetical;
  • Any claim about regulations, policies or prices needs a source and the date it was checked;
  • The old brand name must not appear, and the Chinese can't read like a translation.

If a draft fails, it goes back for a rewrite. Drafts never get published automatically.

What I've learned: getting AI to write a decent article isn't hard. Getting it to stop confidently making things up is. Writing the red lines down as a checklist and handing it to a second AI to enforce is far more reliable than me reviewing every post myself.

4. Personal site: the Chinese briefing now has an English version

My site publishes a Chinese investment research briefing every day. This week I added an automatic English summary. The most important rule in the prompt: only use facts, numbers, dates and ratings that appear in the original. If the original doesn't have it, it doesn't go in.

Two other things while I was at it: the text in buildlog illustrations is now in English, and I launched a fixed-price services page.

(The briefing is just my own research notes. It's not investment advice.)

Three things I learned this week

  1. Clear rules matter more than pretty prompts. A red-line checklist is an asset you can reuse.
  2. Make AIs check each other: keep writing separate from reviewing, and generating separate from acceptance testing.
  3. Let the machines surface the failures. The autoland regression, the compliance review and the broken-link checks all catch problems while I'm asleep.

Do you have a red line of your own when you work with AI? Share it in the comments. I'll pick a few for next week's checklist.