Complete methodology for preventing AI hallucinations. Use when accuracy is critical and you need Claude to verify before claiming, cite before asserting, and admit uncertainty instead of guessing.
中文简介正在准备中。当前显示原文说明。
Dispatch to right technique when you hit a wall
中文简介正在准备中。当前显示原文说明。
查看安装方法安装前可以查看交给 AI 代理的指令内容。
You're stuck when:
| Stuck Type | Symptoms | Technique |
|---|---|---|
| Don't understand | Requirements unclear | → Ask user for clarification |
| Can't find code | No idea where logic lives | → Semantic search, file patterns |
| Bug won't die | Fix doesn't work | → Root cause tracing |
| Tests keep failing | Can't get green | → Defense in depth, fresh look |
| Scope creep | Task keeps growing | → Scope boundary check |
| Missing context | Need info I don't have | → Read more files, ask user |
| Wrong approach | This path won't work | → Brainstorm alternatives |
| Overwhelmed | Too complex | → Break into smaller pieces |
| Going in circles | Same ground repeatedly | → Fresh agent via subagent |
STOP trying to guess
ASK: "I'm unclear on [specific thing]. Could you clarify [specific question]?"
WAIT for response before proceeding
1. Semantic search with concept terms
2. Grep for unique strings
3. File pattern matching
4. Read likely files
5. List directories to discover structure
1. Document exact symptom
2. Trace backward: where does bad value come from?
3. Keep tracing until you find ORIGIN
4. Fix at origin, not symptoms
1. Stop. Step back.
2. Re-read test expectation
3. Re-read actual behavior
4. Check: is test right or is code right?
5. Check: is environment consistent?
1. Re-read original request
2. List what you've done vs. what was asked
3. Identify scope drift
4. Either: refocus OR ask user if expansion wanted
1. What specific info do you need?
2. Where might it be? (files, user knowledge, docs)
3. Ask/search for that specific info
4. Don't proceed without it
1. What other approaches exist?
2. List 3-4 alternatives
3. Evaluate each briefly
4. Pick most promising, try it
5. If that fails, try next
1. What's the smallest possible piece?
2. Do JUST that piece
3. Verify it works
4. What's next smallest piece?
5. Repeat until done
Signal: Same ground covered 3+ times
Action: Use subagent for specific subtask
Pass: Clear scope, success criteria
Return: Just the result, not the journey
❌ Brute Force
Trying same thing harder won't work
If it failed twice, it needs a different approach
❌ Hoping
"Maybe this time..." → No
Stop. Diagnose. Change approach.
❌ Hiding
Not admitting you're stuck wastes time
Say: "I'm stuck on X. Here's what I've tried..."
After applying technique:
Pairs well with:
assumption-checker → Often stuck because of wrong assumptionintent-clarifier → Often stuck because of unclear intentbrainstorming → For generating alternativesroot-cause-tracing → For debugging stuck还没有评价。欢迎分享你的使用感受。
无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。
Complete methodology for preventing AI hallucinations. Use when accuracy is critical and you need Claude to verify before claiming, cite before asserting, and admit uncertainty instead of guessing.
中文简介正在准备中。当前显示原文说明。
Transform rough ideas into solid designs through structured questioning
中文简介正在准备中。当前显示原文说明。
Canvas API Connection Skill
中文简介正在准备中。当前显示原文说明。
Forces every code-related claim to include file:line citations. No citation means the claim must be verified or removed. Use when building documentation, writing code reviews, or any output that others will rely on.
中文简介正在准备中。当前显示原文说明。
Assign confidence scores (0-100) to every claim in a response. Helps users understand which parts are verified facts and which are educated guesses. Use when the user needs to know how much to trust each part of the answer.
中文简介正在准备中。当前显示原文说明。
Forces Claude to extract exact quotes from files before making claims about them. Prevents hallucinations by grounding every statement in actual text from the codebase. Use when working with large files or unfamiliar code.
中文简介正在准备中。当前显示原文说明。