Prompt Rescue Rules

# Prompt Rescue Rules

Prompt Rescue rewrites vague user feedback into actionable prompts.

## Low-Quality Feedback Patterns

- “还是不行”
- “继续”
- “你自己解决”
- “不要出现 bug”
- “重新改”
- “怎么又报错了”
- “still broken”
- “just fix it”
- “continue”
- “fix it yourself”

## Why These Are Inefficient

They do not tell the AI:

- what failed
- what the user saw
- what changed since last time
- whether the previous fix was tested
- what evidence is available
- what should not be changed

This can cause blind edits, broad refactors, repeated mistakes, and longer conversations.

## Rewrite Requirements

A rescued prompt should ask the AI to:

1. Stop broad changes temporarily.
2. Restate the current failure.
3. List the last changed files or actions.
4. Identify likely failure points.
5. Ask for the minimum evidence needed.
6. Avoid unrelated refactors.
7. Define how the next fix will be verified.

## Self-Contained Follow-Up Prompt Rule

The final “你现在最应该问 AI 的一句话” must be self-contained.

It must include:

1. The specific project or feature name.
2. The target that needs verification.
3. The evidence types required, such as code evidence, logs, screenshots, recordings, user
   acceptance, or reproducible steps.
4. A clear instruction that “code was changed” is not enough.
5. A clear stop condition: do not expand features before verification.

If the user did not provide a feature name, use `[具体功能名]` and tell the user to replace it before
sending the prompt.

Weak: