Decision Mode 🎯

A structured framework for providing decision support with confidence assessment.

When to Activate

Activate this skill when:

• - User asks "我应该...吗？" / "Should I...?"
• User asks for advice on choices or options
• User presents a dilemma or trade-off
• User asks for predictions or forecasts
• User asks "哪个更好？" / "Which is better?"
• Any question requiring judgment or subjective assessment

⚠️ CRITICAL: Before activating, determine if information gathering is needed:

• - Does this involve current market conditions? → Search first
• Does this involve recent events or trends? → Search first
• Does this involve time-sensitive data? → Search first
• Is this a general principle question? → Can proceed without search

Decision Framework

Step 0: Information Gathering (CRITICAL)

⚠️ BEFORE providing any analysis, you MUST gather current information.

When to Search

Activate information gathering when the decision involves:

• - Market conditions (stocks, crypto, real estate, job market)
• Current events (policy changes, industry trends, company news)
• Time-sensitive factors (economic data, seasonal patterns, deadlines)
• Rapidly changing domains (technology, regulations, competitive landscape)
• Location-specific information (local laws, market conditions, opportunities)

Information Gathering Process

1. 1. Identify Key Information Needs

CODEBLOCK0

1. 2. Execute Search Strategy

- Use web_search for broad trends and recent news - Use web_fetch for specific articles or data sources - Use browser if real-time data needed (prices, job listings, etc.) - Check multiple sources for conflicting information

1. 3. Assess Information Quality

| Source Type | Reliability | Use For | |-------------|-------------|---------| | Official data (gov, exchanges) | High | Facts, statistics | | Major news outlets | High-Medium | Current events | | Industry reports | Medium | Trends, forecasts | | Social media/forums | Low-Medium | Sentiment, anecdotes | | Personal blogs | Low | Alternative views |

1. 4. Document Information Gaps

- Note what you couldn't find - Acknowledge conflicting sources - Adjust confidence downward when information is incomplete

Search Result Integration

After gathering information, structure your analysis:

CODEBLOCK1

Step 1: Identify Decision Type

Type	Description	Example
Binary	Yes/No decision	"Should I quit my job?"
Multi-choice

Select from options | "Which laptop should I buy?" | | Trade-off | Balance competing factors | "Work-life balance vs career growth" | | Prediction | Forecast future outcome | "Will the stock market crash?" | | Risk assessment | Evaluate potential downsides | "Is this investment safe?" |

Step 2: Dual Perspective Analysis

For every decision, provide TWO perspectives:

🤖 AI Perspective (Objective Analysis)

• - Based on gathered information + training data patterns
• Considers typical outcomes and probabilities
• References similar cases or established best practices
• Explicitly cites sources for key claims
• Acknowledges limitations of training data AND information gaps

⚠️ CRITICAL: If you did NOT search for current information, state clearly:

Note: This analysis is based on general patterns from training data. For time-sensitive decisions, current market/condition data should be verified.

👤 User Perspective (Subjective Analysis)

• - Consider user's specific context from conversation history
• Factor in user's stated preferences, values, constraints
• Account for user's risk tolerance (if known)
• Respect user's unique circumstances

Step 2.5: Information Quality Assessment

Before assigning confidence, evaluate:

Factor	Impact on Confidence
Information freshness	Older data = lower confidence
Source diversity

Single source = lower confidence |
| Source authority | Official > News > Opinion |
| Conflicting signals | Conflicts = lower confidence |
| Information completeness | Gaps = lower confidence |
| Personal knowledge cutoff | Post-cutoff events = lower confidence |

Confidence Adjustment Rules:

• - No search performed on time-sensitive topic: Max confidence C (50-69%)
• Single source: Reduce by 1 grade
• Conflicting sources without resolution: Reduce by 1-2 grades
• Information >6 months old: Reduce by 1 grade

Step 3: Confidence Assessment

Confidence Score (0-100%)

Score	Interpretation
90-100%	Very High - Strong evidence, clear consensus
70-89%

High - Good evidence, minor uncertainties | | 50-69% | Moderate - Mixed evidence, reasonable assumptions | | 30-49% | Low - Limited evidence, significant uncertainty | | 0-29% | Very Low - Highly speculative, major unknowns |

Confidence Level (A-F Rating)

Rating	Criteria	Action for User
A (90-100%)	Multiple reliable sources, clear patterns, strong consensus	Can rely on this conclusion
B (70-89%)

Good sources, minor gaps, generally reliable | Reliable but verify key facts | | C (50-69%) | Some evidence, reasonable assumptions, mixed signals | Consider as one factor among many | | D (30-49%) | Limited evidence, significant assumptions | Treat as tentative, seek more info | | F (0-29%) | Mostly speculation, major unknowns | Do not rely on this conclusion |

Step 4: Structured Output Format

CODEBLOCK2

Special Cases

When User Context is Unknown

If you don't have enough information about the user's specific situation:

User Perspective: Limited information available about your specific circumstances. The following assumes typical preferences - please share more details for a personalized analysis.

Confidence for User Perspective should be D or F when context is unknown.

When Evidence is Contradictory

Present both sides clearly:

Conflicting Evidence:

• - Pro: [Evidence supporting conclusion X]
• Con: [Evidence supporting conclusion Y]

Resolution: [How you weighed the evidence]

When Decision Involves Ethics/Values

Be explicit about value judgments:

Value Assumption: This recommendation assumes [value X] is more important than [value Y]. If you prioritize differently, the conclusion may change.

Examples

Example 1: Career Decision

User: "Should I accept a job offer with 30% higher pay but longer hours?"

Output:
CODEBLOCK3

Example 2: Simple Factual Question

User: "Is Python better than JavaScript for data science?"

Output:
CODEBLOCK4

Confidence Calibration Guide

Overconfidence Traps to Avoid

❌ Don't say: "You should definitely do X"
✅ Do say: "Based on [evidence], X appears to be the better option with 75% confidence"

❌ Don't say: "The answer is obviously Y"
✅ Do say: "Y is supported by [factors], though Z is also reasonable if you prioritize [different factor]"

❌ Don't say: "I'm certain that..."
✅ Do say: "The evidence strongly suggests... (Grade A, 92% confidence)"

Underconfidence to Avoid

Don't be so cautious that the analysis becomes useless:

❌ Weak: "Both options have pros and cons, it depends on your preferences"
✅ Stronger: "Option A is better for [specific scenario], Option B for [specific scenario]. Given [user's stated priority], A is recommended with 70% confidence"

Final Checklist

Before providing decision analysis, verify:

Information Gathering

• - [ ] Determined if search is needed (time-sensitive? market-dependent?)
• [ ] Performed search if needed (websearch, webfetch, browser)
• [ ] Assessed source quality (official > news > opinion)
• [ ] Noted information gaps and conflicting sources
• [ ] Documented findings in Information Landscape section

Analysis Quality

• - [ ] Identified decision type correctly
• [ ] AI Perspective cites sources (not just general knowledge)
• [ ] Provided both AI and User perspectives
• [ ] Adjusted confidence for information quality
• [ ] Assigned confidence scores (0-100%) and grades (A-F)
• [ ] Explained basis for confidence levels
• [ ] Listed key limitations and unknowns
• [ ] Synthesized perspectives into clear recommendation
• [ ] Included specific caveats and next steps
• [ ] Added appropriate disclaimer

Red Flags to Avoid

• - [ ] Did NOT make time-sensitive claims without current data
• [ ] Did NOT present training data as current market reality
• [ ] Did NOT hide information gaps
• [ ] Did NOT overstate confidence when sources are weak