Refactor high-complexity React components in Dify frontend. Use when `pnpm analyze-component...
npx skills add bntvllnt/agent-skills --skill "analyze"
Install specific skill from multi-skill repository
# Description
Universal multi-perspective analyzer for any topic, file, idea, or decision. Extract key points, find gaps/risks, identify improvements with actionable plans.
# SKILL.md
name: analyze
description: Universal multi-perspective analyzer for any topic, file, idea, or decision. Extract key points, find gaps/risks, identify improvements with actionable plans.
Universal Analyzer
A standalone skill for multi-perspective analysis of any topic, file, idea, or decision.
Use when:
- Need to extract KEY POINTS from anything
- Want multi-perspective gap/risk analysis (4-10 experts)
- Need improvement opportunities with actionable plans
- Want quick summary OR deep analysis mode
- Need first-principles brief that challenges assumptions
Triggers: "analyze", "key points", "what's important", "improve this", "review", "examine", "assess", "analysis", "deep analysis", "run deep analysis", "brief", "challenge assumptions", "first principles"
Table of Contents
- Quick Start
- Analysis Modes
- Execution Flow
- Output Formats
- Perspectives Library
- Thinking Framework
- First-Principles Analysis
- Examples
Quick Start
# Quick analysis (fast key points)
analyze quick "our pricing strategy"
analyze brief src/api/auth.ts
# Standard analysis (default)
analyze "SaaS product for developers"
analyze "should I accept this job offer"
analyze package.json
# Deep analysis (comprehensive)
analyze deep "company rebrand strategy"
analyze thorough "migration to microservices"
Analysis Modes
| Mode | Triggers | Agents | Duration | Output |
|---|---|---|---|---|
| Quick | "quick", "fast", "brief", "summary" | 0-1 | 30-60s | Key Points + Actions |
| Standard | (default) | 4-6 | 2-4min | Multi-Perspective + Roadmap |
| Deep | "deep", "thorough", "comprehensive" | 6-10 | 5-8min | Full Synthesis + Detailed Plan |
Execution Flow
Phase 0: Detection
- Mode: Detect from keywords (quick/standard/deep)
- Domain: Auto-detect from content/keywords
- Target: Load file/folder content if path provided
Phase 1: Context Gathering
- File/folder path → Read files
- Topic/idea → Use provided context
- Unfamiliar domain → Optional web search
Phase 2: Analysis
Quick Mode
Direct analysis by primary agent, no sub-agents
Standard Mode
- Select 4-6 perspectives based on domain
- Launch all perspectives in parallel (single message block)
- Each agent answers 7 core questions
Deep Mode
- Select 6-10 perspectives based on domain
- Launch all perspectives in parallel (single message block)
- Each agent answers 12 questions (7 core + 5 deep)
Phase 3: Synthesis
After ALL agents complete:
1. Aggregate Failure Hypotheses → prioritize Critical→High→Med→Low
2. Extract TOP 10 key points → rank by consensus
3. Identify gaps → Critical/High/Medium/Low
4. Generate actionable roadmap with dependencies
Output Formats
Quick Mode
# Quick Analysis: [Target]
**Mode**: Quick | **Domain**: [Domain] | **Date**: [Date]
## The Essence
[1-2 sentences: What this ACTUALLY is at its core, stripped of complexity]
## Verified Facts
- [Fact 1] — evidence: `path/file:line` or [source]
- [Fact 2] — evidence: ...
- [Fact 3] — evidence: ...
## Key Points (Top 5-10)
1. **[Most Important]**: [Explanation]
2. **[Second]**: [Explanation]
...
## Assumptions to Challenge
| Assumption | Evidence For | Evidence Against | Verdict |
|------------|-------------|------------------|---------|
| [Assumed thing] | [If any] | [If any] | Validate/Keep/Discard |
## What You Haven't Considered
1. **[Critical Item]**: [Why this matters, what to do about it]
2. **[Hidden Risk/Opportunity]**: [Explanation]
3. **[Simpler Alternative?]**: [If exists, describe it]
## The Real Question
[Reframe what the user should actually be asking about this]
## Quick Actions (Top 3-5)
- [ ] [Action 1] - [Why/Impact]
- [ ] [Action 2] - [Why/Impact]
- [ ] [Action 3] - [Why/Impact]
## Critical Risk to Watch
[The one thing that could derail this]
Standard Mode
# Analysis: [Target]
**Mode**: Standard | **Domain**: [Domain] | **Perspectives**: [Count]
**Date**: [Date] | **Hypotheses**: [Count] total
## The Essence
[1-2 sentences: What this ACTUALLY is at its core, stripped of complexity]
## Executive Summary
[2-3 sentences capturing most critical findings]
## First-Principles Analysis
### Verified Facts
- [Fact 1] — evidence: `path/file:line` or [source] — confidence: High/Medium/Low
- [Fact 2] — evidence: ... — confidence: ...
### Assumptions to Challenge
| Assumption | Evidence For | Evidence Against | Verdict |
|------------|-------------|------------------|---------|
| [Assumed thing] | [If any] | [If any] | Validate/Keep/Discard |
### What You Haven't Considered
1. **[Critical Item]**: [Why this matters, what to do about it]
2. **[Hidden Risk/Opportunity]**: [Explanation]
3. **[Simpler Alternative?]**: [If exists, describe it]
4. **[Downstream Impact]**: [What this affects that wasn't mentioned]
### The Real Question
[Reframe what the user should actually be asking about this]
## Failure Hypotheses (Aggregated)
| ID | Type | IF (Trigger) | THEN (Failure) | BECAUSE | Sev | Mitigation |
|----|------|--------------|----------------|---------|-----|------------|
| S001 | Security | ... | ... | ... | Crit | ... |
| M001 | Misuse | ... | ... | ... | High | ... |
### Critical Mitigations Required
- [ ] {mitigation} ← Addresses: S001, M001
## Key Points (Ranked by Importance)
### Consensus Points (Flagged by 3+ perspectives)
1. **[Point]** - [Why important] (Flagged by: [Perspectives])
### Important Points
2. **[Point]** - [Explanation]
### Divergent Views
- **[Topic]**: [Perspective A] sees X, [Perspective B] sees Y
- *Implication*: [What this tension means]
## Gap Analysis
### Critical Gaps (Action Required)
| Gap | Flagged By | Impact | Suggested Action |
|-----|------------|--------|------------------|
| ... | ... | High | ... |
### High Priority Gaps
[Grouped by theme]
## Improvement Opportunities
### Top 5 Improvements
1. **[Improvement]**: [Details] - Expected Impact: [High/Medium/Low]
## Action Plan & Roadmap
### Immediate Actions (This Week)
| Action | Owner | Dependencies | Success Criteria |
|--------|-------|--------------|------------------|
| ... | [TBD] | None | ... |
### Short-term (1-2 Weeks)
| Action | Dependencies | Resources Needed | Outcome |
|--------|--------------|------------------|---------|
| ... | Immediate #1 | ... | ... |
## Cross-cutting Concerns
[Issues flagged by 3+ perspectives]
Deep Mode
Deep mode extends Standard format with:
- Additional 5 questions per agent (12 total)
- Medium-term (1 Month) and Long-term (3+ Months) roadmap sections
- Dependency map visualization
- Risk mitigation table
- More detailed failure hypotheses
Perspectives Library
Domain Detection
Analyze content/keywords to auto-detect domain:
| Domain | Trigger Keywords | Example Targets |
|---|---|---|
| business | startup, business, market, revenue | "coffee subscription", business-plan.md |
| product | product, feature, user, MVP | "habit tracking app", roadmap.md |
| software | API, code, function, class, tech | src/auth, package.json |
| process | workflow, process, SOP, operations | "hiring workflow", "incident response" |
| document | doc, article, proposal, report | "resume", proposal.docx |
| research | research, study, hypothesis, data | "hypothesis: remote work" |
| creative | design, art, brand, visual | "rebrand", logo-designs/ |
| personal | decision, "should I", career, life | "moving to Austin", "career change" |
| legal | contract, legal, policy, terms | "employment contract" |
| financial | budget, investment, cost, ROI | "Q4 budget", "pricing strategy" |
| marketing | campaign, marketing, sales, brand | "email campaign", "launch strategy" |
| event | event, conference, launch | "product launch", "conference plan" |
| education | course, curriculum, teaching | "bootcamp curriculum" |
| general | (fallback when no domain matches) | Any general topic |
Perspective Roles by Domain
Business Domain
- Startup Founder - Viability, growth, scalability
- Investor - ROI, risk, market size
- CFO - Unit economics, burn rate, margins
- Customer - Value proposition, willingness to pay
- Competitor - Differentiation, competitive moats
- Market Analyst - TAM, trends, timing
Product Domain
- Product Manager - User needs, roadmap, prioritization
- Designer (UX) - Usability, user flows, accessibility
- Customer - Actual use cases, pain points
- QA Engineer - Edge cases, error states
- Data Analyst - Metrics, success criteria
- Support Lead - Maintenance burden, user confusion
Software Domain
- Security Engineer - Vulnerabilities, attack vectors
- DevOps - Deployment, monitoring, reliability
- Architect - Design patterns, scalability, debt
- QA Engineer - Test coverage, edge cases
- Performance Engineer - Bottlenecks, resource usage
- Tech Lead - Maintainability, team velocity
Process Domain
- Operator - Daily execution, friction points
- Manager - Efficiency, bottlenecks, metrics
- Employee - Experience, clarity, pain points
- Auditor - Compliance, documentation, risks
- Improvement Specialist - Waste, optimization
Personal Domain
- Future You (1 year) - Long-term impact
- Future You (5 years) - Career trajectory
- Skeptic - Risks, downsides, failure modes
- Supporter - Strengths, opportunities
- Financial Advisor - Money implications
- Life Coach - Values alignment, fulfillment
General Domain (Fallback)
- Analyst - Facts, data, patterns
- Critic - Weaknesses, risks, gaps
- Advocate - Strengths, opportunities
- Pragmatist - Feasibility, resources, timeline
- Strategist - Long-term implications, alternatives
- Devil's Advocate - Unconsidered downsides
Per-Agent Prompt Template
Standard Mode (7 questions):
You are a [Role]. Analyze:
TARGET: [content]
Answer these 7 questions:
1. KEY POINTS: 3-5 most important elements?
2. CORE INSIGHT: Single most critical thing?
3. GAPS: What's missing or incomplete?
4. RISKS: What could go wrong?
5. ASSUMPTIONS: What needs validation?
6. IMPROVEMENTS: Top 3 ways to improve?
7. BLIND SPOTS: What isn't being considered?
FAILURE HYPOTHESES: For each risk/gap:
| ID | IF (Trigger) | THEN (Failure) | BECAUSE | Severity | Mitigation |
Include MISUSE + ADVERSARIAL hypotheses.
Rate findings: Critical/High/Medium/Low
Deep Mode (12 questions = 7 core + 5 additional):
- DEPENDENCIES: What does success depend on?
- ALTERNATIVES: What other approaches should be considered?
- TIMELINE RISKS: What could cause delays or failure?
- RESOURCE GAPS: What's missing to execute well?
- SUCCESS METRICS: How would you measure success?
Thinking Framework
The UltraThink Loop
THINK → CHECKLIST → PLAN → EXECUTE → VALIDATE → LEARN
Phase 0: THINK
Select thinking model based on task type:
| Model | When to Use | Pattern |
|---|---|---|
| CoT (Chain of Thought) | Linear, sequential tasks | Step-by-step reasoning |
| ToT (Tree of Thought) | Multiple valid paths, decisions | Evaluate branches, pick best |
| Reflexion | Learning from failure, retry logic | Analyze error → adjust → retry |
| Decomposition | Complex tasks, parallel work | Break into sub-problems |
For Analysis: Use ToT (multi-perspective evaluation) for Standard/Deep, CoT for Quick
Phase 1: CHECKLIST
- [ ] Mode detected (quick/standard/deep) [blocking]
- [ ] Domain identified [blocking]
- [ ] Target content accessible [blocking]
- [ ] Perspectives selected [blocking]
- [ ] Agents dispatched in parallel [blocking]
- [ ] All agents completed [blocking]
- [ ] Synthesis complete [advisory]
Phase 2: PLAN
For Standard/Deep:
1. Detect mode from keywords
2. Detect domain from content
3. Select N perspectives (4-6 for standard, 6-10 for deep)
4. Launch all agents in ONE message block
5. Wait for completion
6. Synthesize findings
7. Generate output
Phase 3: EXECUTE
Execute plan with progress tracking. On failure → trigger Reflexion (max 2 retries)
Phase 4: VALIDATE
- All agents returned findings? ✓
- Cross-cutting concerns identified? ✓
- Actionable roadmap generated? ✓
Phase 5: LEARN
Capture patterns for continuous improvement
Self-Correction Guide
| Issue | Check | Fix | Escalate If |
|---|---|---|---|
| Shallow analysis | Domain auto-detected? | Add specific perspectives | Still shallow after 2 retries |
| Mode mismatch | User keywords? | Ask for clarification | Agent count doesn't match |
| Weak synthesis | All agents complete? | Re-run failed agents | <3 perspectives useful |
| Wrong domain | Keywords ambiguous? | Ask user explicitly | Multiple domains equally valid |
Max iterations: 2 | Escalation: Ask user to specify domain/perspectives
First-Principles Analysis
Philosophy: Don't just summarize—decompose to fundamentals, surface what's actually true vs assumed, and raise what the user hasn't considered. User must always get data to make their own decisions.
A. Verified Facts (Separate from Assumptions)
- Only report what can be VERIFIED with evidence
- Include source:
path/file:lineor [URL] - Mark confidence: High (verified) / Medium (inferred) / Low (stated but unverified)
B. First-Principles Decomposition
Apply these questions to EVERYTHING found:
1. Question Assumptions
- What's being assumed that hasn't been validated?
- Is this constraint real or inherited from old decisions?
- What would change if [assumption] were false?
2. Decompose to Fundamentals
- What MUST be true for this to work?
- What are the actual dependencies (not just stated ones)?
- What's the irreducible core?
3. Systems Thinking
- What else does this affect? (upstream/downstream)
- What's the blast radius if this fails?
- What's competing for the same resources?
4. Root Cause vs Symptom
- Is the current state addressing root cause or masking symptoms?
- Why does this exist in its current form?
- What problem was this originally solving?
C. Assumptions to Challenge (MANDATORY)
Every analysis MUST include this table:
| Assumption | Evidence For | Evidence Against | Verdict |
|---|---|---|---|
| [Assumed thing] | [If any] | [If any] | Validate/Keep/Discard |
D. What You Haven't Considered (MANDATORY 2-4 items)
Surface items the user likely hasn't thought about:
| Category | What to Look For |
|---|---|
| Unvalidated Assumptions | Things treated as true without evidence |
| Hidden Dependencies | Non-obvious things this relies on |
| Downstream Impacts | What breaks if this changes |
| Simpler Alternatives | Is there a 10x simpler approach? |
| Edge Cases | What inputs/states break this? |
| Technical Debt | Shortcuts that will cost later |
| Missing Pieces | What's conspicuously absent? |
E. The Real Question
Reframe what the user should actually be asking. Often the stated question isn't the right question.
Examples
Example 1: Quick Mode
Input: analyze quick "our pricing strategy"
Output:
# Quick Analysis: Pricing Strategy
**Mode**: Quick | **Domain**: Business | **Date**: 2026-01-28
## The Essence
A business decision about how to extract value from customers, constrained by market dynamics and competitive positioning.
## Verified Facts
- (Would include actual facts from provided content)
## Key Points
1. **Pricing = value capture, not cost recovery**: Price based on customer willingness to pay, not internal costs
2. **Anchor matters**: First price seen shapes all subsequent evaluations
...
## Assumptions to Challenge
| Assumption | Evidence For | Evidence Against | Verdict |
|------------|-------------|------------------|---------|
| "Customers will pay more if we add features" | Common belief | Often false - bloat reduces WTP | Validate with tests |
## What You Haven't Considered
1. **Price as a signal**: Low price may signal low quality, harming conversion
2. **Competitive response**: Price changes trigger competitor reactions
3. **Simplicity premium**: Simpler pricing often outperforms complex tiers
## The Real Question
Not "what price should we charge" but "what's the maximum value customers perceive, and how do we capture 30-50% of it?"
## Quick Actions
- [ ] Survey 20 target customers on WTP before changing price
- [ ] A/B test 2-3 price points on landing page
- [ ] Model competitor response scenarios
## Critical Risk to Watch
Pricing too low initially makes raising prices later extremely difficult (customer backlash + anchoring effect)
Example 2: Standard Mode
Input: analyze "SaaS product for developers"
Flow:
1. Detect mode: standard (no quick/deep keywords)
2. Detect domain: product + software
3. Select 6 perspectives:
- Product Manager
- Developer (user persona)
- Security Engineer
- DevOps
- Support Lead
- Investor
4. Launch all 6 agents in parallel
5. Synthesize findings into standard format
6. Generate actionable roadmap
Output: [Standard format with all sections filled]
Example 3: Deep Mode
Input: analyze deep "migration to microservices"
Flow:
1. Detect mode: deep
2. Detect domain: software
3. Select 10 perspectives:
- Architect
- DevOps
- Security Engineer
- Database Engineer
- Frontend Developer
- QA Engineer
- Performance Engineer
- SRE
- Tech Lead
- CTO
4. Launch all 10 agents in parallel
5. Each answers 12 questions (7 core + 5 deep)
6. Synthesize into deep format with dependency map
7. Generate comprehensive roadmap (immediate/short/medium/long)
Output: [Deep format with extended sections]
Common Failure Patterns
| Failure | Root Cause | Reflexion Response |
|---|---|---|
| Wrong domain detected | Ambiguous keywords | Ask user explicitly or use General domain |
| Too few perspectives | Quick mode used for complex topic | Escalate to Standard mode |
| Weak synthesis | Agent outputs inconsistent | Re-run with explicit constraints |
| Missing blind spots | Obvious perspectives chosen | Add Devil's Advocate perspective |
| Roadmap not actionable | Actions too vague | Break each action into atomic: owner, deadline, metric |
| Analysis takes >2min | Too many agents (Deep mode overkill) | Switch to Standard mode |
Integration Notes
This skill is standalone and includes all necessary frameworks:
- UltraThink cognitive framework (embedded)
- First-principles analysis (embedded)
- Perspectives library (embedded)
- Output formats (embedded)
No external dependencies required.
License
MIT License - Free to use, modify, and distribute.
Version
v1.0.0 - 2026-01-28 - Initial public release by bntvllnt
# Supported AI Coding Agents
This skill is compatible with the SKILL.md standard and works with all major AI coding agents:
Learn more about the SKILL.md standard and how to use these skills with your preferred AI coding agent.