@Yesterday
Recent fanfare around Claude Code compels me to write down what I know.
- It is an incredibly intuitive product.
- Leveraging Plan Mode is key. I have a simple heuristic; the more time I spend in plan mode, the better off I am.
- Phase-based updates and planning are good.
- Diagrams are good.
- Building context in a mono repo is good. Context is nothing more than past work and past files, in a literal sense.
- Opus 4.5 was a real unlock; its amazing.
Things I don’t use enough
- Structured permissions for my Claude → it’s dumb that I have to issue permissions for every bash command.
- Custom agents → there are things I can do to create common agents. One idea is to have my Claude Code review my entire history and suggest some options for improvement.
- Structured workflows
- tests
- PR templates
- reviewer/optimizer agents
- Git worktrees
Things I don’t believe in:
- Letting the model run freely (some might say amok) in a Gas Town like approach. I don’t say this from a critical perspective. I just think it depends on having a ton of credits.
- That engineers are obsolete; systems depend on humans understanding them. Systems no one understands are systems no one can fix. Engineers are so important for good design.