investigation-workflow（调查工作流程）

Investigation Workflow Skill

Purpose

This skill provides a systematic 6-phase workflow for investigating and understanding existing systems, codebases, and architectures. Unlike development workflows optimized for implementation, this workflow is optimized for exploration, understanding, and knowledge capture.

When to Use This Skill

Investigation Tasks (use this workflow):

• "Investigate how the authentication system works"
• "Explain the neo4j memory integration"
• "Understand why CI is failing consistently"
• "Analyze the reflection system architecture"
• "Research what hooks are triggered during session start"

Development Tasks (use DEFAULT_WORKFLOW.md instead):

• "Implement OAuth support"
• "Build a new API endpoint"
• "Add feature X"
• "Fix bug Y"

Core Philosophy

Exploration First: Define scope and strategy before diving into code Parallel Deep Dives: Deploy multiple agents simultaneously for efficient information gathering Verification Required: Test understanding through practical application Knowledge Capture: Document findings to prevent repeat investigations

The 6-Phase Investigation Workflow

Phase 1: Scope Definition

Purpose: Define investigation boundaries and success criteria before any exploration.

Tasks:

• FIRST: Identify explicit user requirements - What specific questions must be answered?
• Use prompt-writer agent to clarify investigation scope
• Use ambiguity agent if questions are unclear
• Define what counts as "understanding achieved"
• List specific questions that must be answered
• Set boundaries: What's in scope vs. out of scope
• Estimate investigation depth needed (surface-level vs. deep dive)

Success Criteria:

• Clear list of questions to answer
• Defined scope boundaries
• Measurable success criteria (e.g., "can explain system flow", "can diagram architecture")

Deliverables:

• Investigation scope document with:

- Core questions to answer - Success criteria - Scope boundaries (what's included/excluded) - Estimated depth and timeline

Phase 2: Exploration Strategy

Purpose: Plan which agents to deploy and what to investigate, preventing inefficient random exploration.

Tasks:

• Use architect agent to design exploration strategy
• Use patterns agent to check for similar past investigations
• Identify key areas to explore (code paths, configurations, documentation)
• Select specialized agents for parallel deployment in Phase 3
• Create investigation roadmap with priorities
• Identify potential dead ends to avoid
• Plan verification approach (how to test understanding)

Agent Selection Guidelines:

• For code understanding: analyzer, patterns agents
• For system architecture: architect, api-designer agents
• For performance issues: optimizer, analyzer agents
• For security concerns: security, patterns agents
• For integration flows: integration, database agents

Success Criteria:

• Clear exploration roadmap
• List of agents to deploy in Phase 3
• Prioritized investigation areas

Deliverables:

• Exploration strategy document with:

- Investigation roadmap - Agent deployment plan for Phase 3 - Priority order for exploration - Expected outputs from each exploration

Phase 3: Parallel Deep Dives

Purpose: Deploy multiple exploration agents simultaneously to gather information efficiently.

CRITICAL: This phase uses PARALLEL EXECUTION by default.

Tasks:

• Deploy selected agents in PARALLEL based on Phase 2 strategy
• Common parallel patterns:

- [analyzer(module1), analyzer(module2), analyzer(module3)] - Multiple code areas - [analyzer, patterns, security] - Multiple perspectives on same area - [architect, database, integration] - System architecture exploration

• Each agent explores their assigned area independently
• Collect findings from all parallel explorations
• Identify connections and dependencies between findings
• Note any unexpected discoveries or anomalies

Parallel Agent Examples:

Investigation: "How does the reflection system work?"
→ [analyzer(~/.amplihack/.claude/tools/amplihack/hooks/), patterns(reflection), integration(logging)]

Investigation: "Why is CI failing?"
→ [analyzer(ci-config), patterns(ci-failures), integration(github-actions)]

Investigation: "Understand authentication flow"
→ [analyzer(auth-module), security(auth), patterns(auth), integration(external-auth)]

Success Criteria:

• All planned agents deployed and completed
• Findings from each exploration collected
• Connections between findings identified

Deliverables:

• Findings report with:

- Summary from each parallel exploration - Code paths and flow diagrams - Architectural insights - Unexpected discoveries - Open questions for verification

Phase 4: Verification & Testing

Purpose: Test and validate understanding through practical application.

Tasks:

• Create hypotheses based on Phase 3 findings
• Design practical tests to verify understanding:

- Trace specific code paths manually - Examine logs and outputs - Test edge cases and assumptions - Verify configuration effects

• Run verification tests
• Document what was tested and results
• Identify gaps in understanding
• Refine hypotheses based on test results
• Repeat verification for any unclear areas

Verification Examples:

Understanding: "Authentication uses JWT tokens"
Verification: Trace actual token creation and validation in code

Understanding: "CI fails because of dependency conflict"
Verification: Check CI logs, reproduce locally, verify fix works

Understanding: "Reflection analyzes all user messages"
Verification: Examine reflection logs, trace message processing

Success Criteria:

• All hypotheses tested
• Understanding verified through practical tests
• Gaps in understanding identified and filled

Deliverables:

• Verification report with:

- Tests performed - Results and observations - Confirmed understanding - Remaining gaps or uncertainties

Phase 5: Synthesis

Purpose: Compile findings into coherent explanation that answers original questions.

Tasks:

• Use reviewer agent to check completeness of findings
• Use patterns agent to identify reusable patterns discovered
• Synthesize findings from Phases 3-4 into coherent explanation
• Create visual artifacts (diagrams, flow charts) if helpful
• Answer each question from Phase 1 scope definition
• Identify what worked well vs. what was unexpected
• Note any assumptions or uncertainties remaining
• Prepare clear explanation suitable for user

Synthesis Outputs:

1. Executive Summary: 2-3 sentence answer to main question
2. Detailed Explanation: Complete explanation with supporting evidence
3. Visual Aids: Diagrams showing system flow, architecture, etc.
4. Key Insights: Non-obvious discoveries or patterns
5. Remaining Unknowns: What's still unclear or uncertain

Success Criteria:

• All Phase 1 questions answered
• Explanation is clear and complete
• Findings supported by evidence from verification
• Visual aids clarify complex areas

Deliverables:

• Investigation report with all 5 synthesis outputs
• Ready for knowledge capture in Phase 6

Phase 6: Knowledge Capture

Purpose: Create durable documentation so this investigation never needs to be repeated.

Tasks:

• Store discoveries in memory using store_discovery() from amplihack.memory.discoveries
• Update.claude/context/PATTERNS.md if reusable patterns found
• Create or update relevant documentation files
• Add inline code comments for critical understanding
• Optional: Create GitHub issue for follow-up improvements
• Optional: Update architecture diagrams if needed
• Ensure future investigators can find this knowledge easily

Documentation Guidelines:

## Discovery: [Brief Title]

**Context**: What was investigated and why
**Key Findings**:

- Main insight 1
- Main insight 2
  - **Supporting Evidence**: Links to code, logs, or verification tests
  - **Implications**: How this affects the project
  - **Related Patterns**: Links to similar patterns in PATTERNS.md

Success Criteria:

• Discoveries stored in memory for future reference
• Relevant documentation files updated
• Knowledge is discoverable by future investigators
• No information loss

Deliverables:

• Discoveries stored in memory
• Updated PATTERNS.md (if applicable)
• Updated project documentation
• Optional: GitHub issues for improvements
• Investigation session log in ~/.amplihack/.claude/runtime/logs/

Transitioning to Development Workflow

After investigation completes, if the task requires implementation (not just understanding), transition to DEFAULT_WORKFLOW.md:

1. Resume at Step 4 (Research and Design) with the knowledge gained from investigation
2. Or resume at Step 5 (Implement the Solution) if the investigation already provided clear design guidance
3. Use investigation findings from memory (via get_recent_discoveries()) and session logs to inform design decisions

Example Hybrid Workflow:

User: "/ultrathink investigate how authentication works, then add OAuth support"

Phase 1: Investigation
→ Run INVESTIGATION_WORKFLOW.md (6 phases)
→ Complete understanding of existing auth system
→ Store findings in memory via discoveries adapter

Phase 2: Development
→ Transition to DEFAULT_WORKFLOW.md
→ Resume at Step 4 (Research and Design)
→ Use investigation insights to design OAuth integration
→ Continue through Step 15 (implementation → testing → PR)

When to Transition:

• Investigation reveals implementation is needed
• User explicitly requested both investigation + development
• Follow-up work identified during knowledge capture

Efficiency Targets

Target Efficiency: This workflow targets a 30-40% reduction in message count compared to ad-hoc investigation.

Efficiency Gains Come From:

1. Scope Definition prevents scope creep and wandering
2. Exploration Strategy prevents random unproductive exploration
3. Parallel Deep Dives maximize information gathering speed
4. Verification Phase catches misunderstandings early
5. Synthesis ensures all questions answered
6. Knowledge Capture prevents repeat investigations

Comparison to DEFAULT_WORKFLOW.md

Similarities (Structural Consistency)

Both workflows share core principles:

• Explicit phases with clear deliverables
• Agent-driven execution at each phase
• Quality gates preventing premature progression
• Knowledge capture and documentation
• TodoWrite tracking for progress management

Differences (Investigation vs. Development)

Phase Mapping (For User Familiarity)

Integration with UltraThink

UltraThink Workflow Detection: When /ultrathink is invoked, it automatically detects investigation tasks using keywords and suggests this workflow.

Automatic Workflow Suggestion:

User: "/ultrathink investigate how authentication works"

UltraThink: Detected investigation task. Using INVESTIGATION_WORKFLOW.md
→ Reading workflow from .claude/workflow/INVESTIGATION_WORKFLOW.md
→ Following 6-phase investigation workflow
→ Starting Phase 1: Scope Definition

Customization

To customize this workflow:

1. Edit ~/.amplihack/.claude/workflow/INVESTIGATION_WORKFLOW.md to modify, add, or remove phases
2. Adjust agent deployment strategies for your needs
3. Add project-specific investigation patterns
4. Update efficiency targets based on your metrics

Changes take effect immediately for future investigations.

Success Metrics

Track these metrics to validate workflow effectiveness:

• Message Count: Target 30-40% reduction vs. ad-hoc (to be validated)
• Investigation Time: Track time to completion
• Knowledge Reuse: How often memory retrieval prevents repeat work
• Completeness: Percentage of investigations with full documentation
• User Satisfaction: Clear understanding achieved

Key Principles

• Scope first, explore second - Define boundaries before diving in
• Parallel exploration is key - Deploy multiple agents simultaneously in Phase 3
• Verify understanding - Test your hypotheses in Phase 4
• Capture knowledge - Always store discoveries in memory in Phase 6
• This workflow optimizes for understanding, not implementation

When in doubt about investigation vs. development:

• Investigation: "I need to understand X"
• Development: "I need to build/fix/implement X"

Related Resources

• Source Workflow: ~/.amplihack/.claude/workflow/INVESTIGATION_WORKFLOW.md (complete 436-line specification)
• Knowledge Extraction: Use knowledge-extractor skill after investigations to capture learnings
• Agent Catalog: ~/.amplihack/.claude/agents/CATALOG.md for all available agents
• Pattern Library: ~/.amplihack/.claude/context/PATTERNS.md for reusable investigation patterns