Here's a pro tip regarding Claude Code usage if you're a dev - save your Claude skills and more to git.

This gives you a mechanism to:

  • keep all your setups in sync (personal, work, etc)
  • iteratively develop your setup (skills)
  • blame and revert bad changes
  • and more++

setup

for a simple chill setup:

  1. git init ~/.claude

  2. add the following .gitignore

    # always ignore
cache/
debug/
file-history/
ide/
session-env/
sessions/
shell-snapshots/
telemetry/
usage-data/
backups/
downloads/
history.jsonl
plugins/
agents/**/*.jsonl
projects/**/*.jsonl
projects/**/*.json
projects/**/*.txt

# rm below if machines are similar (os, term setup)
settings.json
statusline-command.sh

# rm below if folder structure across machines is compatible
projects/

# rm below for record-keeping
plans/
todos/

  3. push to a private repo

  4. push and pull from various comps to keep things in sync.

(you can also get fancier and use dotfiles or GNU Stow)

~/.claude tree

claude saves everything to disk in a neat little unobscured .claude folder in your home dir:

~/.claude
|   
|   # main stuff
├── CLAUDE.md
├── skills/
├── commands/
├── agents/
├── hooks/
|
|   
|   # settings
├── settings.json
├── statusline-command.sh
|
|   
|   # project mem
├── projects/
|
|   
|   # session stuff
├── plans/
├── todos/
|
|   
|   # system
├── plugins/  # may have cache or binaries
├── cache/
├── debug/
├── file-history/
├── telemetry/
└── history.jsonl

fin

I abuse Claude like no other, and with this setup I gain the ability to:

  • keep all my setups in sync (personal, work, etc)
  • track/iterate/blame/rollback
    • CLAUDE.md
    • skills
    • hooks
    • agents
    • settings
  • share skills with others
  • see what I did in a day by looking at plans
  • try skills in other harnesses (like Codex) with a simple symlink, since skills are now an open standard

AI is an amazing tool that has augmented the way I code. It seems fitting I treat it like a proper tool in my toolkit.