SoftDeveloper23

I'm building 10 apps in 2026, in public. No team. No funding. Just shipping small software and digital products while sharing the real numbers, the wins, and what breaks. Here is what I am building and how I work.

Giving an AI subagent a vague instruction creates vague results. I give each reviewer four things: exact input, one clear question, what it can edit, and evidence the next person can check. I break down the workflow here: youtu.be/0W_xWWThMUE

GPT-7 already? I looked at the rumors, the public evidence, and what would actually matter for people building with AI. Here's what's known, what's speculation, and what I'd want to see next. www.youtube.com/watch?v=rTMk...

Plot twist: AI might make native mobile development cheaper than “write once, run everywhere.” Shopify is moving back from React Native to Swift and Kotlin because coding agents changed the math. That is a much bigger story than another framework war. shopify.engineering/back-to-native

AI coding gets more reliable when the workflow does not stop at “it works.” I use a simple loop: set a clear goal, reproduce the failure, repair what broke, and check the result again. What does your verification step look like? youtu.be/f5_EiGWvnvs

A test expected 2 warnings. The tool returned 3. The test still passed. The third warning was legitimate. The expected answer was incomplete, and the comparison was too loose to expose it. Here is how an AI reviewer caught the problem: youtu.be/0W_xWWThMUE

AI built the tool. A separate AI code review found the bug. My project-summary tool was counting a checked task as unfinished. This video follows the failure, the parser fix, and the verification loop that came next. youtu.be/f5_EiGWvnvs

A clean SwiftUI build is not proof that the screen works. I am trying to separate three checks when AI coding agents help: build, interaction, and visual comparison. Passing one does not excuse skipping the other two. What usually catches the next problem for you? #SwiftUI

Working from bed today because I’m sick. The plan is not to “push through.” It is one small task, one clear stopping point, and no pretending a low-energy day should work like a normal one. Good systems should survive bad days.

I nearly built a custom redirect today just to make one tracked link look cleaner. Then I tested the actual Bluesky composer. The native card kept the full UTM even after the ugly URL disappeared. What small test recently saved you from building the wrong thing?