madappgang/task-complexity-router

Task Complexity Router

Version: 1.0.0 Purpose: Intelligent task routing to optimal model tiers for cost efficiency and performance Status: Production Ready

Overview

Task complexity routing is the practice of matching tasks to appropriate model tiers based on complexity, urgency, and resource requirements. Instead of using expensive premium models for all tasks, routing directs simple tasks to fast/cheap models and reserves expensive models for complex work.

This skill provides battle-tested patterns for:

• 4-tier routing system (Native Tools → Haiku → Sonnet → Opus)
• Complexity detection heuristics (keyword-based + context-based)
• Cost optimization strategies (save 60-90% on API costs)
• Dynamic tier escalation (upgrade when task stalls or fails)
• Routing integration (works with multi-agent-coordination, quality-gates, proxy-mode)

Well-designed routing can reduce AI costs by 60-90% while maintaining quality, since 70% of tasks can be handled by faster, cheaper models.

Why Task Routing Matters

Cost Comparison (per 1M tokens)

| Model Tier | Model Example | Cost (Input/Output) | Speed | Use Case | |------------|---------------|---------------------|-------|----------| | Tier 0 | Native Tools | $0 | Instant | File operations, searches, formatting | | Tier 1 | Claude Haiku 4.5 | $0.80 / $4.00 | Fast | Simple edits, docs, straightforward tasks | | Tier 2 | Claude Sonnet 4.5 | $3.00 / $15.00 | Moderate | Standard dev, multi-file changes | | Tier 3 | Claude Opus 4.5 | $15.00 / $75.00 | Slower | Architecture, complex debugging, audits |

Example Cost Savings:

Scenario: 100 tasks per day (mix of simple and complex)

Without Routing (all Sonnet):
  100 tasks × 1000 tokens avg × $0.015 = $1.50/day
  Annual: $547.50

With Smart Routing:
  50 tasks (native tools) × $0 = $0
  30 tasks (Haiku) × $0.004 = $0.12
  15 tasks (Sonnet) × $0.015 = $0.22
  5 tasks (Opus) × $0.075 = $0.37
  Total: $0.71/day
  Annual: $259.15

Savings: $288.35/year (52% reduction) for single developer

Performance Benefits

Beyond cost savings, routing improves:

• Speed: Fast models return results in 1-2s vs 10-15s for premium
• Throughput: Process 5x more simple tasks in parallel
• Resource efficiency: Save premium model quota for critical tasks
• User experience: Instant results for simple operations

The 4-Tier Routing System

Tier 0: Native Tools (No LLM)

When to Use:

• File operations (search, rename, move, copy)
• Content search (grep, regex)
• Code formatting (prettier, black, go fmt)
• Git operations (status, log, diff)
• Single-file edits with clear pattern

Indicators:

• Keywords: "find", "search", "format", "rename", "list", "show"
• Patterns: Single regex, exact string replacement, file path operations

Cost: $0 Speed: Instant (< 0.1s)

Examples:

✓ "Find all .tsx files in src/"
✓ "Search for 'TODO' comments"
✓ "Format code with prettier"
✓ "Rename Button.js to Button.tsx"
✓ "Show git status"
✓ "Replace 'oldName' with 'newName' in file.ts"

Implementation:

Task: "Find all TypeScript files"
→ Use Glob tool: *.ts
→ No LLM needed

Task: "Search for API endpoints"
→ Use Grep tool: "app\.(get|post|put|delete)"
→ No LLM needed

Task: "Format all code"
→ Use Bash: bun run format
→ No LLM needed

---

Tier 1: Fast Model (Haiku)

When to Use:

• Simple code changes (add comment, fix typo, rename variable)
• Documentation updates (README, JSDoc, inline comments)
• Straightforward bug fixes (missing import, syntax error)
• Code explanation (what does this function do?)
• Simple test writing (unit test for pure function)

Indicators:

• Keywords: "simple", "basic", "small", "quick", "minor", "add", "fix", "update"
• Scope: Single file, < 50 lines changed
• Complexity: No architectural decisions, clear solution

Cost: ~$0.0004 per task (1000 tokens) Speed: Fast (1-3s response time)

Examples:

✓ "Add JSDoc comment to calculateTotal function"
✓ "Fix typo in error message"
✓ "Rename getUserData to fetchUserData"
✓ "Update README with new installation steps"
✓ "Add missing import statement"
✓ "Write unit test for add(a, b) function"

Anti-Patterns (Don't Use Haiku For):

✗ "Design authentication system" (needs Opus)
✗ "Refactor entire codebase" (needs Sonnet + context)
✗ "Debug complex race condition" (needs Opus)
✗ "Architect database schema" (needs Opus)

---

Tier 2: Standard Model (Sonnet)

When to Use:

• Standard feature implementation (new component, API endpoint)
• Multi-file refactoring (rename class, extract service)
• Integration tasks (connect frontend to backend)
• Moderate bug fixes (logic errors, edge cases)
• Test suites (integration tests, E2E tests)

Indicators:

• Keywords: "implement", "create", "build", "refactor", "integrate", "develop"
• Scope: 2-10 files, 50-500 lines changed
• Complexity: Requires understanding context, moderate problem-solving

Cost: ~$0.003 per task (1000 tokens) Speed: Moderate (5-10s response time)

Examples:

✓ "Implement user profile page with React"
✓ "Create REST API endpoint for /users/:id"
✓ "Refactor authentication logic into AuthService"
✓ "Fix pagination bug in user list"
✓ "Write integration tests for payment flow"
✓ "Add error handling to API calls"

This is the Default Tier: When in doubt, use Sonnet. It handles 70% of standard development tasks well.

---

Tier 3: Premium Model (Opus)

When to Use:

• Architecture decisions (system design, database schema)
• Complex debugging (race conditions, memory leaks, security issues)
• Security audits (vulnerability analysis, threat modeling)
• Performance optimization (algorithm complexity, bottleneck analysis)
• Code review (deep analysis, architectural feedback)
• Critical bug fixes (production outages, data corruption)

Indicators:

• Keywords: "architect", "design", "audit", "complex", "system-wide", "critical", "optimize"
• Scope: System-wide impact, 10+ files, architectural changes
• Complexity: Requires deep reasoning, multiple trade-offs

Cost: ~$0.015 per task (1000 tokens) Speed: Slower (15-30s response time)

Examples:

✓ "Design microservices architecture for e-commerce platform"
✓ "Audit authentication system for security vulnerabilities"
✓ "Debug intermittent race condition in WebSocket handler"
✓ "Optimize algorithm for 1M+ record processing"
✓ "Review entire codebase for architectural issues"
✓ "Design database schema for multi-tenant SaaS"

When to Escalate to Opus:

Task starts in Sonnet, but:
  - Task fails after 2 attempts → Escalate to Opus
  - User explicitly says "this is complex" → Escalate to Opus
  - Implementation reveals architectural issues → Escalate to Opus
  - Performance/security concerns discovered → Escalate to Opus

---

Complexity Detection Heuristics

Keyword-Based Routing

Scoring Algorithm:

Step 1: Extract keywords from user request

Step 2: Score each keyword:
  Tier 0 indicators: +0 points
    - find, search, list, show, format, rename, move, copy, grep

  Tier 1 indicators: +1 point
    - simple, basic, small, quick, minor, add, fix, update, comment

  Tier 2 indicators: +2 points
    - implement, create, build, refactor, integrate, develop, feature

  Tier 3 indicators: +3 points
    - architect, design, audit, complex, system-wide, critical, optimize

Step 3: Calculate total score
  Score 0: Use native tools (Tier 0)
  Score 1-2: Use Haiku (Tier 1)
  Score 3-5: Use Sonnet (Tier 2)
  Score 6+: Use Opus (Tier 3)

Step 4: Apply context modifiers (next section)

Example Scoring:

Request: "Add simple comment to function"
Keywords: "add" (+1), "simple" (+1), "comment" (+1)
Score: 3 → Sonnet (Tier 2)
Context Modifier: Single file → -1 → Score 2 → Haiku (Tier 1)

Request: "Implement user authentication"
Keywords: "implement" (+2), "authentication" (+3)
Score: 5 → Sonnet (Tier 2)

Request: "Design microservices architecture"
Keywords: "design" (+3), "microservices" (+3), "architecture" (+3)
Score: 9 → Opus (Tier 3)

Request: "Find all TODO comments"
Keywords: "find" (+0)
Score: 0 → Native tools (Tier 0)

---

Context-Based Routing

File Count Modifier:

Files affected (from user context or codebase analysis):
  1 file → -1 tier (simpler)
  2-5 files → No modifier
  6-10 files → +0 tier (standard)
  11+ files → +1 tier (complex)

Example:
  Task: "Refactor authentication" (base Tier 2)
  Context: Affects 15 files
  Modifier: +1 tier → Opus (Tier 3)

Code Complexity Modifier:

Indicators of complexity (increase tier):
  - Async/await patterns (+1)
  - Error handling required (+1)
  - Database transactions (+1)
  - Security implications (+2)
  - Performance critical (+2)
  - System-wide impact (+2)

Example:
  Task: "Fix login bug" (base Tier 2)
  Context: Security implications (+2)
  Final: Opus (Tier 3)

User Context Modifier:

User explicitly signals complexity:
  "This is simple" → -1 tier
  "This is complex" → +1 tier
  "Be careful" → +1 tier
  "Quick task" → -1 tier
  "Critical" → +1 tier

Example:
  Task: "Update README" (base Tier 1)
  User: "Be careful, this affects onboarding"
  Modifier: +1 tier → Sonnet (Tier 2)

---

Override Patterns

User-Specified Model:

User explicitly requests tier:
  "Use Haiku to add comment"
  → Override routing, use Haiku (Tier 1)

  "Use Opus to review this"
  → Override routing, use Opus (Tier 3)

Priority: User override > Routing algorithm

Fallback Strategy:

When routing is uncertain:
  - Score is borderline (e.g., 2.5 between tiers)
  - Mixed signals (simple keywords, complex context)
  - No clear indicators

Action: Default to Sonnet (Tier 2)
  - Safe choice for most tasks
  - Not too expensive ($0.003 vs $0.015)
  - Good quality for standard work
  - Can escalate to Opus if needed

Emergency Escalation:

Task fails at current tier:
  Attempt 1: Use routed tier (e.g., Haiku)
  Attempt 2: Same tier, different approach
  Attempt 3: Escalate +1 tier (e.g., Sonnet)
  Attempt 4: Escalate to Opus (highest tier)

Example:
  Task: "Fix subtle bug" → Haiku (Tier 1)
  Result: Fails to identify root cause
  → Retry with Haiku (different prompt)
  Result: Still fails
  → Escalate to Sonnet (Tier 2)
  Result: Identifies bug, fixes it ✓

---

Cost Optimization Patterns

Cost-Benefit Analysis

When to Upgrade Tier:

Upgrade from Tier 1 (Haiku) to Tier 2 (Sonnet):
  Cost increase: $0.002 (0.5x more)

  Upgrade when:
    - Task failed 2+ times at Tier 1
    - Task requires multi-file context
    - Risk of incorrect solution is high
    - Time spent debugging > cost savings

Upgrade from Tier 2 (Sonnet) to Tier 3 (Opus):
  Cost increase: $0.012 (5x more)

  Upgrade when:
    - Task has critical security/performance implications
    - Architecture decisions needed
    - Task failed 2+ times at Tier 2
    - Complex reasoning required (trade-offs, edge cases)

When to Downgrade Tier:

Downgrade from Tier 2 (Sonnet) to Tier 1 (Haiku):
  Cost savings: $0.002 (50% reduction)

  Downgrade when:
    - Subtask is simpler than parent task
    - Clear, straightforward solution exists
    - Single file, < 50 lines changed
    - No architectural decisions needed

Example:
  Main task: "Implement user profile" → Sonnet (Tier 2)
  Subtask 1: "Add JSDoc to ProfileCard" → Haiku (Tier 1)
  Subtask 2: "Write unit test for formatDate" → Haiku (Tier 1)
  Subtask 3: "Integrate with API" → Sonnet (Tier 2)

---

Cost Tracking Integration

Track Costs Per Task:

Task: "Implement user authentication"
Model: Claude Sonnet 4.5
Tokens: 1500 input, 3000 output
Cost: (1500 × $0.003 / 1000) + (3000 × $0.015 / 1000)
     = $0.0045 + $0.045
     = $0.0495

Log to performance tracking:
  {
    "task": "Implement user authentication",
    "tier": 2,
    "model": "claude-sonnet-4-5",
    "tokens_in": 1500,
    "tokens_out": 3000,
    "cost": 0.0495,
    "duration_seconds": 8,
    "success": true
  }

Aggregate Cost Metrics:

Daily Cost Report:
  Tier 0 (Native): 45 tasks, $0.00
  Tier 1 (Haiku): 30 tasks, $0.12 ($0.004 avg)
  Tier 2 (Sonnet): 20 tasks, $0.60 ($0.030 avg)
  Tier 3 (Opus): 5 tasks, $0.75 ($0.150 avg)

  Total: 100 tasks, $1.47
  Routing efficiency: 75% tasks in Tier 0-1 (cheap)

Cost Optimization Insights:

Analysis: Top 5 expensive tasks
  1. "Design database schema" - Opus - $0.25
  2. "Security audit" - Opus - $0.20
  3. "Refactor auth system" - Sonnet - $0.08
  4. "Implement profile page" - Sonnet - $0.06
  5. "Debug race condition" - Opus - $0.15

Recommendation:
  - Tasks 1-2 correctly used Opus (architecture + security)
  - Task 3 could have been split into subtasks (Haiku for simple parts)
  - Task 5 correctly escalated to Opus after Sonnet failed

---

Integration with Orchestration Plugin

Routing + Multi-Agent Coordination

Pattern: Task Delegation with Routing

Workflow: Implement feature with multiple agents

Step 1: Architect designs system (Tier 3 - Opus)
  Task: "Design authentication architecture"
  Model: Opus (complex, architectural)
  Output: Architecture plan

Step 2: Split into subtasks with routing
  Subtask 1: "Implement JWT service" → Sonnet (Tier 2)
  Subtask 2: "Add JSDoc comments" → Haiku (Tier 1)
  Subtask 3: "Write unit tests" → Haiku (Tier 1)
  Subtask 4: "Integrate with API" → Sonnet (Tier 2)

Step 3: Delegate to appropriate agents
  backend-developer (Sonnet): Subtasks 1, 4
  documenter (Haiku): Subtask 2
  test-writer (Haiku): Subtask 3

Cost:
  Without routing: 4 tasks × Sonnet = $0.12
  With routing: 1×Opus ($0.015) + 2×Sonnet ($0.06) + 2×Haiku ($0.008) = $0.083
  Savings: 31%

---

Routing + Multi-Model Validation

Pattern: Tiered Review for Cost Efficiency

Use Case: Code review with budget constraints

Step 1: Fast pre-review (Tier 1 - Haiku)
  Task: "Check for obvious issues: syntax, imports, formatting"
  Model: Haiku (fast, cheap)
  Output: Found 3 obvious issues (missing import, typo, formatting)

Step 2: Fix obvious issues
  Developer fixes issues found by Haiku

Step 3: Deep review (Tier 3 - Opus)
  Task: "Security audit and architectural review"
  Model: Opus (complex, critical)
  Output: Found 2 security issues

Cost:
  Without tiering: Opus review all issues = $0.025
  With tiering: Haiku pre-review ($0.001) + Opus deep review ($0.015) = $0.016
  Savings: 36%
  Benefit: Haiku caught simple issues fast, Opus focused on complex issues

---

Routing + Quality Gates

Pattern: Escalation on Failure

Workflow: Test-driven development with escalation

Iteration 1: Run tests (Tier 2 - Sonnet)
  Task: "Analyze test failures and fix"
  Model: Sonnet
  Result: Fixed 8/10 failures
  Cost: $0.003

Iteration 2: Re-run tests (Tier 2 - Sonnet)
  Task: "Fix remaining 2 failures"
  Model: Sonnet
  Result: Still failing (complex race condition)
  Cost: $0.003

Iteration 3: Escalate to Opus (Tier 3)
  Task: "Debug complex race condition"
  Model: Opus (escalated due to failure)
  Result: Identified root cause, fixed ✓
  Cost: $0.015

Total Cost: $0.021
Without escalation: 3 × Opus = $0.045 (214% more expensive)
With escalation: Try cheaper first, upgrade only when needed

---

Routing + Proxy Mode

Pattern: External Model Routing

Use Case: Use external fast models for simple tasks

Task: "Add comments to functions"
Complexity: Simple (Tier 1)

Option 1: Claude Haiku 4.5 (OpenRouter)
  Cost: $0.004
  Speed: 2s

Option 2: DeepSeek Coder (OpenRouter)
  Cost: $0.001 (75% cheaper than Haiku)
  Speed: 3s

Option 3: Grok Code Fast (OpenRouter)
  Cost: $0.002 (50% cheaper than Haiku)
  Speed: 1s (fastest)

Routing Decision:
  If speed priority: Use Grok Code Fast
  If cost priority: Use DeepSeek Coder
  If balance: Use Claude Haiku (best quality/cost/speed)

Implementation:
  Task: task-executor
    Model: grok
    Prompt: "Add JSDoc comments to all functions in UserService.ts"
    claudish: grok

---

Best Practices

Do:

• ✅ Use native tools (Tier 0) for file operations and searches (instant, free)
• ✅ Start with cheaper tiers and escalate only when needed
• ✅ Track costs per task for optimization insights
• ✅ Set max cost budgets per workflow (e.g., $0.10 per feature)
• ✅ Split complex tasks into simpler subtasks (route each separately)
• ✅ Use Haiku for documentation, comments, simple tests
• ✅ Use Sonnet as default for standard development
• ✅ Reserve Opus for architecture, security, complex debugging
• ✅ Escalate tier after 2 failures at current tier
• ✅ Document routing decisions for team alignment

Don't:

• ❌ Use Opus for every task (waste money, slower)
• ❌ Use Haiku for complex tasks (will fail, waste time)
• ❌ Ignore user context (security/critical tasks need higher tier)
• ❌ Skip cost tracking (can't optimize what you don't measure)
• ❌ Over-optimize for cost (quality matters more than pennies)
• ❌ Use LLM when native tools work (search, format, rename)
• ❌ Downgrade tier for critical tasks (security, production bugs)
• ❌ Route without fallback strategy (always have Sonnet default)

Performance Benchmarks:

Response Time (avg):
  Native Tools: < 0.1s
  Haiku: 1-3s
  Sonnet: 5-10s
  Opus: 15-30s

Cost Efficiency (per 1000 tasks):
  All Opus: $150 (baseline)
  All Sonnet: $30 (80% savings)
  Smart Routing: $7 (95% savings vs Opus, 77% vs Sonnet)

Quality (task success rate):
  Native Tools: 100% (deterministic)
  Haiku: 85% (simple tasks only)
  Sonnet: 95% (most tasks)
  Opus: 98% (all tasks)

---

Examples

Example 1: Analyzing User Request and Determining Tier

Scenario: User requests feature implementation

User Request:

"Add user authentication to the app"

Routing Analysis:

Step 1: Extract keywords
  Keywords: "add" (+1), "authentication" (+3)
  Base Score: 4 → Sonnet (Tier 2)

Step 2: Analyze context
  Scope: Multiple files (routes, middleware, database)
  Complexity: Security implications (+2)
  Adjusted Score: 6 → Opus (Tier 3)

Step 3: Routing decision
  Tier: 3 (Opus)
  Reason: Security-critical, architectural decision
  Cost: ~$0.015 per subtask

Step 4: Task breakdown with routing
  Main task → Opus:
    "Design authentication architecture"
    Output: Architecture plan with security considerations

  Subtasks → Route separately:
    1. "Implement JWT service" → Sonnet (Tier 2)
    2. "Add password hashing" → Sonnet (Tier 2, security)
    3. "Create login endpoint" → Sonnet (Tier 2)
    4. "Add JSDoc comments" → Haiku (Tier 1)
    5. "Write unit tests" → Haiku (Tier 1)
    6. "Write integration tests" → Sonnet (Tier 2)

Total Cost:
  1 × Opus ($0.015) + 4 × Sonnet ($0.012) + 2 × Haiku ($0.008)
  = $0.035

Without routing (all Opus):
  7 × Opus = $0.105
  Savings: $0.07 (67% reduction)

---

Example 2: Multi-Task Workflow with Mixed Tiers

Scenario: Daily development workflow

Tasks:

Task 1: "Find all TypeScript files in src/"
  Keywords: "find" (+0)
  Routing: Tier 0 (Native Tools)
  Tool: Glob "src/**/*.ts"
  Cost: $0
  Time: < 0.1s

Task 2: "Format all code"
  Keywords: "format" (+0)
  Routing: Tier 0 (Native Tools)
  Tool: Bash "bun run format"
  Cost: $0
  Time: 2s

Task 3: "Add JSDoc to UserService"
  Keywords: "add" (+1), "simple" (implied)
  Routing: Tier 1 (Haiku)
  Model: claude-haiku-4-5
  Cost: $0.004
  Time: 2s

Task 4: "Implement user profile page"
  Keywords: "implement" (+2)
  Context: Multiple files (component, styles, API)
  Routing: Tier 2 (Sonnet)
  Model: claude-sonnet-4-5
  Cost: $0.030
  Time: 8s

Task 5: "Fix pagination bug"
  Keywords: "fix" (+1), "bug" (+2)
  Context: Requires debugging
  Routing: Tier 2 (Sonnet)
  Model: claude-sonnet-4-5
  Cost: $0.025
  Time: 7s

Task 6: "Security audit of auth system"
  Keywords: "security" (+3), "audit" (+3)
  Context: Critical, system-wide
  Routing: Tier 3 (Opus)
  Model: claude-opus-4-5
  Cost: $0.150
  Time: 25s

Daily Summary:
  Total Tasks: 6
  Native Tools: 2 tasks, $0, 2s
  Haiku: 1 task, $0.004, 2s
  Sonnet: 2 tasks, $0.055, 15s
  Opus: 1 task, $0.150, 25s

  Total Cost: $0.209
  Total Time: 44s

Without Routing (all Sonnet):
  6 tasks × $0.030 avg = $0.180 (but slower, less optimized)

Without Routing (all Opus):
  6 tasks × $0.150 avg = $0.900 (4.3× more expensive)

---

Example 3: Cost Comparison Showing Savings from Routing

Scenario: Weekly development for small team (3 developers)

Task Distribution:

Week 1: 150 total tasks

Task Breakdown (with smart routing):
  Tier 0 (Native): 50 tasks (file ops, searches, formatting)
    Cost: $0
    Time: 50 × 1s = 50s

  Tier 1 (Haiku): 45 tasks (comments, docs, simple fixes)
    Cost: 45 × $0.004 = $0.18
    Time: 45 × 2s = 90s

  Tier 2 (Sonnet): 45 tasks (features, refactors, integrations)
    Cost: 45 × $0.030 = $1.35
    Time: 45 × 8s = 360s

  Tier 3 (Opus): 10 tasks (architecture, security, complex bugs)
    Cost: 10 × $0.150 = $1.50
    Time: 10 × 25s = 250s

Total with Routing:
  Cost: $3.03
  Time: 750s (12.5 min)
  Tasks completed: 150

Cost Comparison:

Strategy 1: All Opus (premium everywhere)
  150 tasks × $0.150 = $22.50
  Difference: +$19.47 (7.4× more expensive)

Strategy 2: All Sonnet (standard everywhere)
  150 tasks × $0.030 = $4.50
  Difference: +$1.47 (1.5× more expensive)

Strategy 3: Random 50/50 Haiku/Sonnet (no routing logic)
  75 × $0.004 + 75 × $0.030 = $2.55
  Quality issues: 15 tasks failed (Haiku used for complex)
  Re-work: 15 × $0.030 = $0.45
  Total: $3.00 (similar cost, but failures and delays)

Strategy 4: Smart Routing (this approach)
  Cost: $3.03
  Quality: High (each task matched to appropriate tier)
  Speed: Optimal (fast models for simple, premium for complex)

Annual Savings (3 developers × 52 weeks):
  vs All Opus: 52 × $19.47 = $1,012.44
  vs All Sonnet: 52 × $1.47 = $76.44
  vs Random: 52 × $0.45 (re-work) = $23.40

ROI:
  Save $1,000+ annually vs premium-everywhere approach
  Save $76 annually vs standard-everywhere approach
  Better quality than random routing

---

Troubleshooting

Problem: Task routed to Haiku, but failed

Cause: Task more complex than keywords suggested

Solution: Escalate to next tier (Sonnet) and retry

❌ Wrong:
  Task fails with Haiku
  → Give up or ask user for help

✅ Correct:
  Task fails with Haiku (Attempt 1)
  → Retry with different prompt (Attempt 2)
  → Still fails? Escalate to Sonnet (Attempt 3)
  → Fixed ✓

---

Problem: Too many tasks routed to Opus (high costs)

Cause: Overly aggressive routing, not splitting tasks

Solution: Break complex tasks into simpler subtasks

❌ Wrong:
  Task: "Implement user authentication" → Opus
  Cost: $0.150

✅ Correct:
  Main task: "Design auth architecture" → Opus ($0.015)
  Subtask 1: "Implement JWT service" → Sonnet ($0.030)
  Subtask 2: "Add JSDoc comments" → Haiku ($0.004)
  Subtask 3: "Write tests" → Haiku ($0.004)
  Total: $0.053 (65% savings)

---

Problem: Routing adds complexity, slows down workflow

Cause: Over-engineering routing for small projects

Solution: Use simple heuristics for small teams

Simple Routing Rule (for small projects):
  - File operations → Native tools
  - Single file < 50 lines → Haiku
  - Everything else → Sonnet
  - User says "complex" → Opus

No need for complex scoring algorithms for < 50 tasks/day.

---

Problem: Can't track costs (no visibility)

Cause: Missing cost tracking integration

Solution: Add logging to track cost per task

After each task:
  Log to performance-tracking.json:
  {
    "task": "Implement profile page",
    "tier": 2,
    "model": "claude-sonnet-4-5",
    "tokens_in": 1200,
    "tokens_out": 2800,
    "cost": 0.0456,
    "duration_seconds": 8,
    "success": true,
    "timestamp": "2026-01-28T10:30:00Z"
  }

Weekly aggregation:
  Generate cost report with breakdown by tier

---

Summary

Task complexity routing optimizes AI workflows by:

• 4-tier system (Native → Haiku → Sonnet → Opus)
• Complexity detection (keyword-based + context-based scoring)
• Cost optimization (60-90% savings vs premium-everywhere)
• Dynamic escalation (upgrade tier when task fails)
• Quality maintenance (match task complexity to model capability)

Key Takeaways:

1. Use native tools first (free, instant)
2. Start cheap, escalate as needed (Haiku → Sonnet → Opus)
3. Track costs to optimize (measure, analyze, improve)
4. Split complex tasks (route subtasks separately)
5. Reserve premium for critical (security, architecture, debugging)

Master routing and you'll deliver faster results at a fraction of the cost.

---

Inspired By:

• Cost optimization patterns from production AI workflows
• Multi-tier routing systems (CDN, load balancers, database replication)
• Google's multi-model serving strategy (PaLM vs Gemini routing)
• AWS Lambda tiered pricing (pay for what you use)
• Performance budgets in frontend development (optimize critical path)