Skip to main content

​
1. Start with a Clear, Specific Goal

The quality of your high-level objective directly impacts the agent’s performance.
  • Bad Goal: “Fix my project.”
  • Good Goal: “In the ./src directory, refactor all Python files to use the logging module instead of print statements. Ensure all changes pass the existing pytest suite.”
A specific goal reduces the chance of the agent misinterpreting the task and performing unintended actions.

​
2. Operate in a Version-Controlled Environment

This is your safety net. Always run Agent Mode on a project that is under version control (e.g., Git).
  • Commit Before Starting: Ensure you have a clean working directory.
  • Review Changes: Use git diff to see exactly what the agent modified.
  • Revert if Needed: If the agent messes up, you can instantly revert to the previous state.

​
3. Supervise the First Run

Before letting the agent run fully autonomously on a large task, supervise its initial steps.
  • Ask for a Plan: Before giving the go-ahead, ask Chibi: “What is your plan to achieve this?”.
  • Check the “Thoughts”: If SHOW_LLM_THOUGHTS is enabled, read the agent’s internal reasoning to ensure it understands the context correctly.
  • Monitor: Keep an eye on the output to ensure it’s not going down a rabbit hole.
Return to: Agent Mode Introduction