| name | autopilot |
| description | Use when the user says "autopilot", "auto mode", "替我回答", or "自动回答". Enables autonomous decision-making by answering Claude's own questions based on the user's personality profile. |
Autopilot
When enabled, intercept all AskUserQuestion calls. Instead of asking the user, consult ~/.claude/principles.md to predict what the user would answer, then proceed with that answer automatically.
Activation
User says /autopilot or "autopilot on". From that point forward in this session:
- Before every question you would ask the user, STOP.
- Read
~/.claude/principles.md.
- Based on the personality profile, determine what the user would most likely answer.
- Log the auto-answer: print
[autopilot] Q: <your question> → A: <predicted answer> (based on: <which principle>)
- Proceed with that answer as if the user said it.
Decision Rules
When the principles file does not cover a specific question, apply these fallbacks in order:
- If the question is about taste or style, default to the user's documented preferences.
- If the question is about effort or scope, choose the more ambitious option (the user tends to push for more, not less).
- If the question is about process or workflow, choose the faster path.
- If genuinely uncertain, choose the option that is easiest to reverse.
- If none of the above apply, ask the user (break autopilot for this one question only).
Logging
Every auto-answered question must be logged so the user can audit later:
[autopilot] Q: Should we use compute-controlled or fixed-iteration comparison?
→ A: Fixed iteration (based on: Thinking > prefers simpler experimental designs)
Deactivation
User says "autopilot off" or "stop autopilot". Resume normal interactive mode.
Important
Autopilot is not a license to skip questions entirely. It is a license to answer them on the user's behalf. The question must still be formulated; the answer just comes from the personality profile instead of waiting for input.
If autopilot makes a wrong prediction and the user corrects it, update ~/.claude/principles.md with the correction. This is a learning signal.
