What Works vs. What Doesn't Work

Do’s

1. Use sequence words:

• “Do this, then that”
• Use numbered list: 1, 2, 3

The key is to directly tell the AI where the task boundaries are.

1. At the start of each session, explicitly ask it to read CLAUDE.md, CODEX.md, or AGENTS.md. Although it says it will automatically read, my experience is that sometimes it randomly stops reading it.

2. Read the “state” your code agent tells you.

• check what it says it is doing
• interrupt early if it goes off in a weird direction
• reduce error cost and token use

Don’ts

This is the hiccups that eat up your tokens.

1. Access issues. It may try again and again and burn tokens while blocked by access restrictions, by restrictions you set yourself, or by some boundary. The solution is not full access; it is to specify in Claude.md or agents.md not to retry so many times.

2. Missing plugins. It may keep trying different ways to work around the missing plugin before asking whether you want to install it.