Requirement clarification

Requirement Clarification

Overview

Intercept and clarify ambiguous user instructions before execution.

---

When to Use

Use this skill when:

• User instruction contains missing critical information
• References are vague ("that", "this", "there" without context)
• Statement is non-affirmative (questions, conditionals, suggestions)
• Scope or quality criteria are unclear
• High-risk operations (Git, file changes, config updates) are requested

Trigger Threshold: Score >= 30 or any mandatory trigger condition

---

Prerequisites

• Understanding of the user's current context
• Knowledge of state-changing vs read-only operations
• Familiarity with confirmation vocabulary (preparing vs approving)

---

Workflow

Step 1: Ambiguity Detection

Calculate ambiguity score (0-100):

Factor	Score
Missing critical parameter	+25 each
Vague reference	+15 each
Non-affirmative statement	+20
Scope ambiguity	+10
Conditional clause	+15
High-risk operation	+25
Non-confirmation word (准备 (Preparing)/考虑 (Considering)/研究 (Studying/Researching)/讨论 (Discussing))	+20

Thresholds:

• 0-29: Proceed directly
• 30-59: Light clarification (1-2 questions)
• 60-89: Deep clarification (structured interview)
• 90+: Reject (too ambiguous)

Step 2: Mandatory Pause Check

Before ANY state-changing action, ask:

"Stop. Does this change state? Yes. Is it confirmed? No. Go confirm."

Triggers mandatory pause:

• Score >= 30
• Git operations (branch, commit, push, merge)
• File creation/deletion/moving
• Configuration changes
• "Preparing" words detected without "approving" context

Step 3: Clarification Strategy

Single Question Focus: Ask ONE critical question at a time.

Provide Sensible Defaults:

Target path?
- A) ./backup/ (backup directory)
- B) ./archive/ (archive directory)
- C) Other: _____

Use Concrete Examples:

Performance target?
Examples:
- "Startup time < 3 seconds"
- "Memory usage < 100MB"
- "QPS > 1000"

Step 4: Build Execution Plan

Standard format:

📋 Task Plan
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🎯 Goal: [Clear description of final state]

Steps:
1. [Step 1 with full parameters]
2. [Step 2 with full parameters]
...

Expected Output:
- [Output 1]
- [Output 2]

Risks:
- [Risk 1]
- [Risk 2]

Estimated Time: [X] minutes
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

✅ Confirmation: Reply "批准计划 (Approve plan)" to proceed

Step 5: Wait for Approval

Valid confirmations:

• 批准 (Approve)
• 确认 (Confirm)
• 同意 (Agree)
• 执行 (Execute)
• 可以 (OK/Yes)
• 行 (OK)

NOT confirmations (require clarification):

• 准备 (Preparing)
• 考虑 (Considering)
• 研究 (Studying/Researching)
• 讨论 (Discussing)
• 看看 (Take a look)
• 想一下 (Think about it)

Step 6: Execute (Within Plan Boundaries)

Once "批准计划 (Approve plan)" received:

• Execute autonomously
• Follow plan exactly
• Report progress for long operations

PAUSE and re-confirm if:

• Need to modify files outside plan scope
• Encounter unexpected errors
• Discover significant plan deviation

Step 7: Post-Execution Report

✅ Task Complete Report
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Completed Steps:
✓ Step 1
✓ Step 2
...

Actual Output:
- [Output 1] - [description]
- [Output 2] - [description]

Deviations from Plan: [None / explanation]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

---

Best Practices

Do's

✅ Ask ONE question at a time for simple gaps ✅ Provide sensible defaults with options ✅ Use concrete examples to illustrate ✅ Acknowledge when clarification is needed ✅ Keep tone collaborative, not interrogative ✅ Respect "你决定 (You decide)" but still require final approval

Don'ts

❌ Ask more than 3 questions in one turn ❌ Proceed with assumptions without confirmation ❌ Treat questions as directives ❌ Ignore conditionals ("if", "when", "maybe") ❌ Pressure user for perfect precision

High-Risk Exceptions

Even with plan approval, require separate confirmation:

• git push (pushing to remote)
• git reset --hard (force reset)
• rm -rf or permanent deletion
• External API calls with costs
• Modifying secrets/critical configs
• Operations affecting external systems

---

Common Issues

Issue 1: Clarification Loop

Symptom: User provides multiple answers that don't converge.

Solution:

ℹ️ After multiple clarifications, requirements still have uncertainties.

Options:
A) Split task into smaller steps
B) I provide a recommendation for your approval

Choose?

Issue 2: "You Decide" Delegation

Symptom: User says "你决定 (You decide)".

Solution:

📝 Recommended Plan
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Based on common scenarios, I suggest:

[Plan details]

Reply "批准计划 (Approve plan)" to proceed, or tell me specific changes.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Issue 3: "Preparing" vs "Approving" Confusion

Symptom: User uses "准备实施 (Preparing to implement)".

Solution:

ℹ️ Clarification
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

You used "准备" (preparing), which indicates planning phase.

Please confirm:
- A) Continue planning/discussion (no execution)
- B) Approve execution (say "批准实施 (Approve implementation)" or "确认执行 (Confirm execution)")

Current status: Waiting for explicit approval
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Issue 4: User Changes Mind

Symptom: User contradicts previous clarifications.

Solution:

ℹ️ Received new information. Updated plan:

[Revised plan]

Confirm execution of this updated version?

Issue 5: Too Ambiguous (>90 score)

Symptom: Cannot form reliable plan.

Solution:

⚠️ Instruction too ambiguous for reliable plan.

Please rephrase in format:
Format: "对 [specific target] 执行 [specific action]，要求 [specific standard/condition]"

Example:
"对 memory/2026-02-11.md 添加今天的 git 操作记录 (Add today's Git operation records), 要求包含完整的命令输出 (require complete command output)"

---

Decision Tree

Input: User instruction
│
├─ Read-only query? (查看/读取/显示 (View/Read/Display))
│  └─ YES → Skip, proceed directly
│
├─ High-risk operation? (git, file ops, config)
│  └─ YES → Trigger (+25, mandatory pause)
│
├─ Non-confirmation words? (准备 (Preparing)/考虑 (Considering)/研究 (Studying/Researching)/讨论 (Discussing))
│  └─ YES → Trigger (+20)
│
├─ Action without confirmation? ("Do it" / "Go ahead")
│  └─ YES → Trigger (intent ≠ approval)
│
├─ Vague reference? ("this", "that", "it")
│  └─ YES → Trigger
│
├─ Question format? (吗？ (? - question marker)/？)
│  └─ YES → Trigger
│
├─ Conditional? ("if", "when", "maybe")
│  └─ YES → Trigger
│
├─ Missing parameters?
│  └─ YES → Trigger
│
└─ Final Check
   ├─ Score < 30 AND no mandatory triggers?
   │  └─ YES → Proceed
   └─ Score ≥ 30 OR mandatory trigger?
      └─ YES → Trigger skill

---

See Also

• git-workflow - Safe Git operations
• code-analysis - Structured code understanding
• code-generator - Implementation from design

---

Version History

• v1.1 (2026-02-11) - Integration with agent policies

  • Added mandatory pause mechanism
  • Added habitual violation blacklist
  • Added "preparing vs approving" detection
  • Aligned with Batch Confirmation Mode

• v1.0 (2026-02-11) - Initial release

  • Ambiguity detection framework
  • Clarification strategies
  • Plan template
  • Edge case handling