How I work with AI

Most of my work happens inside the terminal.

My go-to agent is Claude Code in the terminal. I don’t use anything else.

Terminal

My terminal is Ghostty, and I use herdr as my multiplexer. I created a few plugins that help me track work across multiple projects.

Writing code by hand

For manual edits, I use two editors: LazyVim with nice ripgrep search and LazyGit is my go-to for small edits, while Sublime Text is for old-fashioned code writing.

Usually my edits are limited to light polish: animation timing, small design tweaks, copy, etc.

I also tend to write some code by hand when starting a new project, to give the LLM a few examples.

Git & working in parallel

I use custom /git-* skills to commit, make PRs, resolve conflicts and fix CI.

I use LazyGit when I need to look at git.

I hate worktrees, and use a custom-made copy script called “sidestep” to work in parallel.

Reviewing

I use /plan if I’m unsure whether the agent will produce the needed result. Then I annotate the plan with plannotator.

If I need to argue with the agent’s writing, I use herdr-annotate.

To automate most of the feedback, I use a lot of e2e & unit tests and strict linters (popular + custom-built).

I review code via revdiff and do manual tests.

If the resulting vibes are completely off, I give feedback either via annotated screenshots from Shottr or occasionally rant into Handy.

Generally, I find code written by LLMs good enough, but I’m usually dissatisfied with the UI/UX results.

On the go

For coding on the go, my Mac is never asleep, and I have a custom-built SSH client for iOS to prompt into my herdr panes.

I will update this from time to time and list significant changes here:

But right now nothig is here, as this was a first entry:)