Six Months In: What Actually Changed In How I Build

Six months is enough to see the patterns. I am not writing a retrospective today, just noting what shifted from January to July. The changes are smaller than I expected and more durable than I thought.

Planning Before Code

In January I opened Claude Code and asked for features directly. In July I write a plan file first, review it, then ask the agent to execute. The extra ten minutes up front saves an hour of reversals. This is now automatic.

Smaller Commits

January commits were 500-1000 lines and covered a feature. July commits are 50-200 lines and cover a task. Per-task commits let me revert narrowly and bisect quickly. Git history is now documentation, not just backup.

Structured Returns

Subagents return JSON dictionaries, not prose. Orchestrators branch on specific fields. The shift from free-form to structured handoffs was the single biggest reliability win in my agent workflows.

Less Tooling

My January stack had eleven services. My July stack has eight. Each drop was a conscious audit, not attrition. Fewer moving parts means fewer failure modes and shorter onboarding for anyone (including future me) who touches the code.

More Writing

One daily article. Every day. That habit started in April and did not break. Writing forced me to understand topics I thought I already knew. Half the articles discovered a bug or a bad assumption in my own system while I was writing them.

The Pattern Underneath

All of these are forms of the same thing: more deliberation, smaller units, clearer contracts. None of it is glamorous. It is the plumbing of sustained output. The AI tools got faster over the six months too, but the habit changes mattered more.

What Did Not Stick

Fancy productivity systems. Time-blocking apps. A complicated notes hierarchy. I tried several. None survived contact with real work. The simple habits won every comparison.

Half the year remains. The goal is not another set of changes. It is consistency on the ones that are working.