Notes on working with AI
Clean up after Claude, on purpose
AI-built projects collect duplicates, thin wrappers and temporary fixes. A separate cleanup pass that reports first keeps them manageable.
Three weeks in
After three weeks of building a project with Claude you usually find the same things: two functions doing the same job, a wrapper around a single line, a "temporary" fix from week one that everything now leans on, and files nobody opens. None of it breaks the project. All of it makes the next change slower and riskier.
Why it piles up
Each session solves the problem in front of it. The model adds what the task needs and rarely goes back to remove what an earlier session left behind. It also can't see the shortcut it took last Tuesday unless you show it. A person feels the mess and eventually tidies up. Claude only feels it if you ask.
That points to the fix: a separate cleanup pass, not a smarter build prompt. Give the pass one job, which is to find things, not to change them.
What a good pass looks for
Four categories cover most of it: duplicated logic, thin wrappers that add nothing, temporary workarounds that outlived the problem, and files nobody uses. For each finding, Claude should say what it is, why it thinks it can go, and what could break. You approve item by item, and the tests run after each removal.
It isn't only about code
Finance. A model workbook Claude has patched for a month: duplicate tabs, hardcoded numbers sitting on top of formulas, sheets nobody reads. The same four categories apply.
HR. A Claude Project with fifteen instruction files that now contradict each other. Claude lists the conflicts, you choose which rule wins.
Marketing. A campaign folder with five versions of the same brief. Claude lists them, you name the current one, and the rest gets archived.
A note on ready-made skills
Community skills built for exactly this kind of cleanup are starting to appear. Treat them like any code you download: read what the skill tells Claude to do before you install it, and run it on a branch or a copy first.