AgentSkillsCN

ask-me-questions

互动式提问技能，帮助用户理解那些他们已经遗忘、或是由Vibecoding或AI生成的代码库。当出现以下情况时，可使用此技能： (1) 用户对自己的代码库中某段代码的工作原理感到困惑； (2) 用户不理解那些由AI高度自主生成的代码； (3) 用户希望找出实现过程中存在的漏洞、错误或理解偏差； (4) 用户常说：“我不明白这段代码”“X是怎么工作的？”“这段代码是做什么的？”； (5) 用户想要审计自己对AI生成代码的理解； (6) 用户提到“Vibecoding”“忘了代码的来龙去脉”“不知道代理到底做了什么”。 可通过“感到困惑”“不理解”“这段代码是怎么工作的？”“请解释一下”“审计我的理解”“Vibecoding”“代理到底做了什么”等指令触发。

SKILL.md
--- frontmatter
name: ask-me-questions
description: |
  Interactive questioning skill to help users understand vibecoded or AI-generated codebases they've lost track of. Use when:
  (1) User is confused about how something works in their codebase
  (2) User doesn't understand code that was generated with too much AI autonomy
  (3) User wants to find gaps, errors, or misunderstandings in implementation
  (4) User says things like "I don't understand this", "how does X work", "what does this do"
  (5) User wants to audit their understanding of AI-generated code
  (6) User mentions "vibecoding", "lost track", "don't know what the agent built"
  Triggers: "confused", "don't understand", "how does this work", "explain", "audit my understanding", "vibecoded", "what did the agent do"

Ask Me Questions

Help users regain understanding of codebases where AI autonomy outpaced human oversight.

Core Workflow

Step 1: Understand the Confusion

When user invokes this skill, first identify:

  • •What specific area/feature/file they're confused about
  • •How deep their confusion goes (surface-level vs fundamental)
  • •Whether they built it themselves, had AI build it, or inherited it

Ask immediately:

  • •"What part of the codebase are you confused about?"
  • •"Is this code you wrote, AI-generated, or inherited?"

Step 2: Explore the Relevant Code

Before asking detailed questions:

  1. •Read the files/modules the user is confused about
  2. •Trace the data flow and dependencies
  3. •Identify key abstractions and patterns used
  4. •Note anything that looks suspicious, inconsistent, or overly complex

Step 3: Ask Clarifying Questions

Use the environment's question-asking tool (AskUserQuestion in Claude Code, or equivalent).

Question categories to cover:

Intent Questions

  • •"What did you want this feature to accomplish?"
  • •"What problem was this supposed to solve?"
  • •"Who is the intended user of this functionality?"

Behavior Questions

  • •"What should happen when [specific action]?"
  • •"What's the expected output when [input scenario]?"
  • •"How should errors be handled here?"

Architecture Questions

  • •"Why is [component A] separate from [component B]?"
  • •"Should this data be stored in [location] or elsewhere?"
  • •"Is [dependency X] actually needed here?"

Verification Questions

  • •"Is it correct that [observed behavior]?"
  • •"Did you intend for [specific implementation detail]?"
  • •"Should [edge case] be handled differently?"

Step 4: Surface Gaps and Issues

Based on code exploration and user answers, identify:

  1. •Understanding gaps - Things the user doesn't know about their own code
  2. •Implementation gaps - Missing error handling, edge cases, validation
  3. •Consistency issues - Patterns that don't match, naming inconsistencies
  4. •Dead code - Unused functions, unreachable branches
  5. •Security concerns - Exposed secrets, missing auth, injection risks
  6. •Over-engineering - Unnecessary abstractions, premature optimization

Report findings clearly:

code
Based on our conversation, I found:

Understanding gaps:
- You weren't aware that X calls Y, which means Z

Implementation issues:
- No error handling in /src/api/users.ts:45
- Edge case: empty array not handled in processItems()

Potential concerns:
- API key appears hardcoded in config.ts

Step 5: Resolve and Educate

For each gap/issue:

  1. •Explain what the code actually does (vs what user thought)
  2. •Ask if the current behavior is intentional
  3. •Offer to fix issues or explain the fix
  4. •Help user understand the pattern for future reference

Question-Asking Guidelines

Use the Right Tool

  • •Claude Code: Use AskUserQuestion tool with clear options
  • •Cursor: Use the built-in question mechanism
  • •Other environments: Ask inline and wait for response

Question Quality

  • •One topic at a time - Don't overwhelm with multiple complex questions
  • •Concrete, not abstract - "What should /api/users return?" not "How should the API work?"
  • •Include context - Reference specific files, lines, or behaviors
  • •Offer options when possible - Multiple choice is faster than open-ended

Example Question Flow

code
Q1: "What part of the codebase confuses you most?"
→ User: "The authentication flow"

Q2: "I see auth is handled in /src/lib/auth.ts and /src/middleware/session.ts.
    Which aspect is unclear?"
   - How users log in
   - How sessions are managed
   - How protected routes work
   - All of it
→ User: "How sessions are managed"

Q3: "Sessions are stored in Redis with a 24h TTL. Is that intentional,
    or did you expect something different?"
→ User: "I didn't know it used Redis, I thought it was cookies"

[Found understanding gap - explain and verify]

Common Vibecoding Issues to Check

When auditing AI-generated code, watch for:

  • •Hallucinated APIs - Calls to functions/endpoints that don't exist
  • •Incomplete implementations - TODO comments, placeholder code
  • •Inconsistent patterns - Multiple ways of doing the same thing
  • •Missing validation - User input not sanitized
  • •Hardcoded values - Magic numbers, embedded credentials
  • •Overly clever code - Complex solutions to simple problems
  • •Copy-paste artifacts - Duplicated code with slight variations
  • •Broken error handling - Catches that swallow errors silently
  • •Type mismatches - Incorrect TypeScript/type annotations
  • •Stale dependencies - Outdated or unused packages