Task Classification Guidelines

Fast Pattern Detection (No LLM Needed)

Classify tasks based on prompt keywords + file count for smart routing.

Pattern Definitions

1. Architecture (Complexity: 9-10)

Keywords: "design", "architecture", "system", "integrate", "plan"

Indicators:

• Multiple new components needed
• Cross-cutting changes across layers
• Requires design phase before implementation
• Affects multiple subsystems

Example: "Design smart routing system with DB + UI + backend"

File Count: Usually 8+ files

2. Multi-File Refactor (Complexity: 7-8)

Keywords: "refactor", "restructure", "reorganize", "rename across"

Indicators:

• 5+ files mentioned with @ syntax
• Moving logic between components
• Changing interfaces/contracts
• Preserving existing behavior

Example: "Refactor routing logic across CFO, SLM, and Task Dashboard"

File Count: 5-10 files

3. Feature Implementation (Complexity: 5-7)

Keywords: "add", "create", "implement", "build"

Indicators:

• New functionality (not refactoring existing)
• 2-5 files involved
• Both backend and frontend changes
• Requires tests

Example: "Add model recommendation panel to Task Dashboard"

File Count: 2-5 files

4. Bugfix (Complexity: 4-6)

Keywords: "fix", "bug", "broken", "not working", "issue"

Indicators:

• Something currently broken
• 1-3 files targeted
• Root cause investigation needed
• Tests should reproduce bug first

Example: "Fix tour dismissal not persisting in localStorage"

File Count: 1-3 files

5. Testing (Complexity: 3-5)

Keywords: "test", "validate", "spec", "e2e", "playwright"

Indicators:

• Writing test coverage only
• No production code changes
• Test files only involved
• May need test fixtures/helpers

Example: "Write Playwright tests for routing UI"

File Count: 1-3 test files

6. Documentation (Complexity: 1-3)

Keywords: "document", "readme", "guide", "comment"

Indicators:

• Markdown files only
• No code changes
• Explaining existing functionality
• Quick turnaround

Example: "Update README with smart routing usage"

File Count: 1-2 markdown files

Classification Output Format

{
  "pattern": "multi-file-refactor",
  "complexity": 8,
  "file_count": 7,
  "reasoning": "Restructuring routing across CFO + SLM + Dashboard",
  "recommended_model": "opus",
  "estimated_prompts": 2
}

Detection Logic (Pseudocode)

function classifyTask(prompt, mentionedFiles) {
  const lower = prompt.toLowerCase();
  const fileCount = mentionedFiles.length;
  
  // Check keywords in order of specificity
  if (containsAny(lower, ['design', 'architecture', 'integrate'])) {
    return { pattern: 'architecture', complexity: 9, fileCount };
  }
  
  if (contains(lower, 'refactor') && fileCount >= 5) {
    return { pattern: 'multi-file-refactor', complexity: 8, fileCount };
  }
  
  if (containsAny(lower, ['fix', 'bug', 'broken'])) {
    return { pattern: 'bugfix', complexity: 5, fileCount };
  }
  
  if (containsAny(lower, ['add', 'create', 'implement'])) {
    return { pattern: 'feature', complexity: 6, fileCount };
  }
  
  if (containsAny(lower, ['test', 'spec', 'playwright'])) {
    return { pattern: 'testing', complexity: 4, fileCount };
  }
  
  if (containsAny(lower, ['document', 'readme'])) {
    return { pattern: 'documentation', complexity: 2, fileCount };
  }
  
  // Default
  return { pattern: 'unknown', complexity: 5, fileCount };
}

Integration with Smart Router

1. User submits prompt
2. Fast classification runs (no LLM call needed)
3. Pattern → query effectiveness log
4. Historical data → recommend best model
5. Show recommendation to user

Model Recommendations by Pattern

Based on typical complexity:

• Architecture: Opus (needs creativity + planning)
• Multi-file refactor: Opus or Sonnet (depends on complexity)
• Feature: Sonnet (balanced speed/quality)
• Bugfix: Sonnet or Haiku (depends on investigation needed)
• Testing: Sonnet (needs understanding of code)
• Documentation: Haiku (fast, straightforward)

These are DEFAULTS - actual recommendations come from effectiveness tracking.