Claude Code: what I actually use it for

I use Claude Code for hobby projects and long refactors. At work I still live in an IDE agent most of the day. Those are different jobs. This page is the map for the posts where I learned that the hard way.

What it is good at

The CLI forces you to state the task before it touches files. That sounds like friction. It is the point. I get better diffs when I write the goal first, then let it run.

It finds bugs I walk past — race conditions, missing edge cases, half-finished error paths. Test generation is the biggest time save: write tests, run them, fix failures in a loop. I wrote up the March 2025 tradeoffs in Claude Code strengths and weaknesses. The product moved since then. The shape of the tradeoff did not: great for exploration and multi-file work, weak when you need tight editor context and type-aware autocomplete.

Native install matters less than people think once it is on your machine. How you steer it matters more.

CLAUDE.md and project memory

Models forget your repo's weird rules. You should not. Documenting project quirks in CLAUDE.md is the habit that paid off most after building this site with the agent. One file of "we use X not Y" beats re-explaining monorepo layout every session.

The plugin marketplace is the next layer: reusable skills instead of pasting the same prompt. I care about that more than web/iOS shells of the same agent — see Claude Code on web and iOS if you want the product surface, not the workflow.

Cost is part of the workflow

Cheap models are only cheap if the harness does not waste turns. GLM is cheap; steering it is not. Budget pooling is how I stop a long agent session from becoming a surprise bill. If you only care about list price, you will pick the wrong tool.

Haiku has a place for short tasks. I wrote Claude Haiku 4.5 when the price/speed ratio made sense for grunt work. I still review everything.

Lint has to keep up

AI writes faster than ESLint type-aware rules can run. Biome for AI workflows is the fix I landed on: validate the flood of new code without waiting 30 seconds for a type graph. Slow feedback trains you to skip checks. Fast feedback keeps the agent honest.

What built this blog

Parts of this site were generated and refined under Claude Code. Part 1 and part 2 are the build diary. The stack notes sit in why I built a static blog and the Cloudflare hosting map.

Cursor vs Claude Code

At work, editor context wins. Open files, local types, inline accept/reject. On personal projects, the conversation loop wins. I do not try to force one tool into both jobs. Copilot's agent mode is another competitor in that space — the upgrade note is thin because I did not switch primary tools over it.

What I track next

Plugins that survive more than one repo. Cost controls that do not require a spreadsheet. Whether native app surfaces change how I prompt, or just where I read the output.

Dated posts hang off this page. If a new Claude Code release only changes the UI chrome, I might not write. If it changes how I structure a monorepo session, I will.