Problem Solver

Problem Solver Skill

Description

Sophisticated methodology for diagnosing, researching, and solving any problem through systematic investigation, hypothesis testing, and iterative refinement.

Activation

Auto-activates when:

• Something "not working", "broken", "failing", "unexpected"
• Error messages or stack traces appear
• User asks "why", "how to fix", "what's wrong"
• Debugging or troubleshooting needed

Core Philosophy

Don't guess. Investigate. Prove. Verify.

Methodology: DIQRS-ITVF

Phase 1: DISCOVER

D - Diagnose

• What exactly is the problem? (verbatim errors)
• What's expected vs actual behavior?
• When did it start? What changed recently?
• Is it reproducible? Under what conditions?

I - Intersect

• Cross-reference multiple data sources
• Compare logs, configs, code, and behavior
• Find where symptoms converge
• Identify common threads

Q - Question

Formulate precise questions:

• "Why does X happen when Y?"
• "What controls Z behavior?"
• "Where is this value coming from?"
• Challenge assumptions with questions

Phase 2: RESEARCH

R - Research

• Read all relevant logs and configs
• Search codebase for related patterns
• Check documentation and changelogs
• Look for similar issues (GitHub, forums)

S - Synthesize

• Combine findings into coherent picture
• Identify root cause vs symptoms
• Map the causal chain
• Formulate hypotheses ranked by likelihood

Phase 3: ITERATE

I - Iterate

• Test hypotheses one at a time
• Start with most likely cause
• Make minimal changes
• Document each attempt and result

T - Test

• Verify hypothesis with concrete test
• Reproduce the problem first
• Apply fix and test again
• Check for regressions

V - Verify

• Confirm fix addresses root cause
• Test edge cases
• Ensure no side effects
• Get user confirmation

F - Finalize

• Document the solution
• Update relevant configs/code
• Add safeguards if appropriate
• Share learnings (update CLAUDE.md, create hooks)

Diagnostic Toolkit

Log Analysis

# Claude Code decision audit
tail -100 ~/.claude/logs/decision-audit.jsonl | jq -s 'group_by(.decision) | map({decision: .[0].decision, count: length})'

# Recent errors
grep -i "error\|fail\|denied" ~/.claude/logs/*.json | tail -20

# Session context
cat ~/.claude/logs/session_start.json | jq '.context'

Config Validation

# Validate JSON syntax
cat ~/.claude/settings.local.json | jq . > /dev/null && echo "Valid" || echo "Invalid JSON"

# Check effective settings
cat ~/.claude/settings.local.json | jq '.permissions'

# Find all settings files
find ~/.claude -name "settings*.json" -o -name "*.local.json" 2>/dev/null

System State

# Environment
env | grep -i claude
which claude && claude --version

# Processes
ps aux | grep -E "claude|node|python" | grep -v grep

# Network
lsof -i -P | grep LISTEN | head -10

Problem Pattern Library

Sophisticated Techniques

Multi-Source Triangulation

Don't trust single source. Verify with:

1. Logs (what happened)
2. Config (what should happen)
3. Code (what's implemented)
4. Behavior (what actually happens)

Hypothesis Ranking

Score each hypothesis:

• Evidence strength: How much supports it?
• Explanatory power: Does it explain all symptoms?
• Falsifiability: Can we disprove it easily?
• Simplicity: Occam's razor

Bisection Debugging

When cause unclear:

1. Identify working state and broken state
2. Find midpoint (time, commit, config change)
3. Test midpoint
4. Narrow range by half
5. Repeat until root cause found

Rubber Duck Protocol

Before asking for help:

1. State the problem out loud
2. Explain what you've tried
3. Describe expected vs actual
4. Often solution becomes obvious

Anti-Patterns

• Shotgun debugging: Changing random things hoping something works
• Symptom fixing: Treating symptoms, not root cause
• Assumption blindness: "It can't be X" without checking
• Documentation trust: Assuming docs match reality
• Single hypothesis: Not considering alternatives
• Skipping verification: Assuming fix worked without testing