What Works vs. What Doesn't Work

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. Use high tier models (e.g. GPT 5.6 Sol) with high effort level for simple tasks. It not only burns tokens, but also makes it unnecessarily slow and fill up answers with irrelevant content. My rule of thumbs:
    • If it is a clearly articulated task, use Luna. The answer it gives are structured and concise.
    • If it is a verbal task, be it a writing task or a reasoning task, use Sol. It is more creative and can reason better.
    • For comparison, one session of Luna uses less than 1% of your 5-hour limit, while one session of Sol - medium ability can use up to 20% and with high ability it can take up your entire 5-hour limit. Read more here: Saving Tokens in My Process →

  2. 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.

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