Basics
- Custom Claude commands in personal use
- Define custom commands by creating markdown files in .claude/commands/ (e.g., update_docs.md). use /commit /pr /worktree
- You can paste images with ctrl+v or drag and drop them
- add links and it will search
- Queued messages. You can queue messages that will be added to the context. I am not completely sure how it uses these. More black-boxy that I like, so I don't personally use it. But it is there.
- Interrupt Early, Interrupt Often (press esc)
- double esc lets you edit previous message
Core Workflows
- My prefered workflow is to write stories in a docs/todo.md and run my todo custom command. Spec based engineering.
- Worktrees and pros and cons
- Slot machine. A new way of looking at code. Code is cheap, ideas are forever. Working on a POC, let it run with different ideas and pick what you like.
Context Management
- Context rot. Compacting vs MD based management. Yes, it is all about dealing with finicky cache to tell a make a non determisric machine do work for us.
- Context is everything, use /clear often to load context. /compact + auto compact are terrible
- Context is the most important thing. Managing your context at the right time and doing the compaction on your own is the best way to manage
- Keep Claude.md as clean as you can. You can have your personal claude.md in ~/.claude/claude.md
- Passing all the logs into one place like dev.log makes it easier to feed claude errors to keep it fixing things. Unified logging people are using shoreman. Ask it to tail it
Session Management
- --continue to continue sessions. you can also jump to other sessions
- --resume for old sessions
Tools & Integrations
- Claude code sdk exists, but you can't use it through claude pro/max.
- MCP is redudandant. You can just create bash scripts and expose it to them. setup gh and let claude know about it
- Gh and GitHub. Using CLI tools ripgrep/fd/ etc. Let it run it Ccusage. Claude works in 5 hour blocks .starting from the first message you sent. Different plans get different volume of usage and there is variance based on the amount of people using it.
Magic Words
- Think Deeply -> Ultrathink
- Headless mode: I still haven't explored this. fanning. pipelining
- You can run up to 10 parallel sub-agents
Review flows?
Personal software + workflows: anything nagging you, you can just build it into your workflow. Add a script in bash,bun,etc and throw it into your /bin/. Now you have it everywhere.