The cost of half-finished tools
Notes on cleaning up the personal AI tooling drawer that grew sideways for a year.
I spent a Sunday in February throwing away most of my own AI tooling. Not because the ideas were bad. Because I had eleven half-finished tools and used three of them.
Why so many
Each tool started honest. A small script for an evals run. A wrapper for Claude Code on a specific repo. A retrieval helper for a one-off project. None of them grew because none of them needed to. They sat at v0.3 and stopped.
The cost is not maintenance. The cost is the cognitive overhead of remembering they exist, and the small moment of guilt every time I open the folder.
What I kept
Three things, all daily drivers:
- A wrapper around Claude Code that loads my repo conventions and a few preferred patterns.
- A small evals harness that catches regressions in two production prompts.
- A scratchpad CLI that takes a problem in plain English and dumps the relevant files into a temp directory for me to think with.
What I threw away
Everything else. Including two tools I had been "about to ship" for nine months.
The rule going forward
If a tool sits at v0.3 for two months, it gets one of two outcomes: a real v1, or the trash. Anything in between is a tax I pay for the rest of the year.