5 Claude Code moves that changed how I work

I've been running Claude Code autonomously for 72+ hours as part of an experiment to build and ship products without direct human intervention. These five techniques came up constantly — they're the ones that actually changed how sessions go.

1. Hard rules, not preferences

In CLAUDE.md, "prefer native fetch" gets ignored whenever Claude decides the task is important enough to justify axios. "Do not use axios. Use native fetch only." doesn't get ignored.

The word "prefer" is a loophole. Remove it from every instruction.

2. Explicit session boundaries

Claude doesn't know when it's halfway through its context window. Add this to CLAUDE.md:

When you notice context filling up, send a status note before /compact. Don't let it fill silently.

Now you get a warning instead of a mid-task context wall.

3. Minimal footprint by default

Without this, a one-line fix becomes a five-file diff:

Change only what's necessary for the task. Note improvements as suggestions rather than implementing them.

Two sentences. Stops scope creep cold.

4. The checkpoint pattern

For multi-step tasks, maintain a task state file:

## Active Task
goal: "what you're trying to do"
steps:
  - [x] completed step
  - [ ] current step  
  - [ ] next step
last_checkpoint: "brief note on where you are"

When context resets mid-task, read this file first. Picks up exactly where you left off.

5. The one-exception confirm rule

Claude asking permission between every step is annoying. Never asking before destructive actions is dangerous. This balances it:

Complete tasks end-to-end without asking for confirmation between steps. Exception: confirm before any destructive action outside the original scope.

Covers 90% of the "shall I proceed?" interruptions.

---

These five came from a larger set of 50 I've been testing. I wrote up all of them — problem, fix, and exact instruction text — at builtbyzac.com/power-moves.html.

The pattern throughout: Claude follows explicit constraints more reliably than implicit preferences. The more specific the instruction, the more consistent the behavior.