Clear when the unit is done, not when the session feels long
A tenth of a growing number is still a growing number. The cache slows the climb, it never flattens it, so finishing something is the trigger.
Writing · 4 articles
Articles
Measured notes on shipping mobile apps with AI agents. What held up, what broke, and the numbers behind both. Newest first.
2026
Latest
Your requirements live in a chat window
Six rules from fifteen real OpenSpec changes for writing project docs an agent can't misread, including the one that says a proposal is allowed to fail.
Two hooks that hold every time, and what breaks them
Two Claude Code hooks that block bad Swift before it ships, the exit-code contract behind them, and a subshell bug that let one through anyway.
What survives of prompt engineering
I taught prompt writing twice, then measured where the gain was. Three techniques survived agentic work; the rest was decoration.
Why your Claude Code sessions burn out so fast
The limits are not what is emptying your budget. Context hygiene is, and a handful of habits cuts effective usage in half.
Notes
A Claude Code statusline that shows what's left
The default statusline hides the one number worth watching. Here is the prompt I used to replace it, and the line in it that decides what comes back.