Task Delegation

Task Delegation Skill

Purpose: Guide cost-effective delegation of straightforward tasks to Haiku model (12x cheaper than Sonnet 4.5) while reserving Sonnet for complex reasoning.

When to Use This Skill

Use this skill when you're about to start a task and need to decide:

1. Should I do this directly with Sonnet?
2. Should I delegate this to Haiku?
3. What's the cost-benefit tradeoff?

Model Cost Comparison (January 2026)

Cost multiplier: Haiku is 12x cheaper than Sonnet 4.5

Decision Matrix

✅ ALWAYS Delegate to Haiku

Test-related:

• Updating test assertions (e.g., assertEquals(0.9, x) → assertTrue(x >= 0.9))
• Adding test cases following existing patterns
• Expanding test coverage with edge cases
• Creating integration tests from examples

Refactoring:

• Renaming variables/methods/classes
• Extracting methods
• Applying simple patterns (HashSet → Set.of())
• Converting for-loops to streams (when pattern is clear)

Documentation:

• Updating README, CLAUDE.md, or skill files
• Adding code comments
• Generating JavaDoc
• Creating changelog entries

Bug fixes:

• Fixing NPE with clear root cause
• Correcting off-by-one errors
• Fixing regex patterns (with examples)
• Adding null checks

Code hygiene:

• Adding logging statements
• Removing System.out.println
• Formatting code
• Fixing import statements

❌ NEVER Delegate to Haiku

Architecture:

• Designing new handlers
• Creating similarity calculators from scratch
• Deciding on metadata structure
• API design decisions

Complex debugging:

• Circular initialization issues
• Race conditions
• Flaky tests (investigation required)
• Cross-handler pattern matching bugs
• Memory leaks

Exploration:

• "Investigate why..." questions
• Understanding codebase architecture
• Finding root cause of ambiguous issues
• Performance profiling

First-time patterns:

• Creating the first test for a new calculator
• Establishing a new code pattern
• Writing handlers without examples

Security-sensitive:

• Input validation
• SQL injection prevention
• XSS prevention
• Authentication/authorization logic

Delegation Template

When delegating to Haiku, use this proven template:

Task(
  subagent_type="general-purpose",
  model="haiku",
  prompt=`You are working on <branch-name> for <PR-number>.

## Context
[1-2 sentences explaining what needs to be done]

## Tasks
1. [Specific task with file path]
2. [Specific task with expected outcome]
3. Run tests: mvn test -Dtest=<TestClass>
4. Verify all tests pass

## Success Criteria
- [Concrete measurable outcome]
- [Test count or assertion format]
- [No compilation errors]

## Reference Examples
Look at these files for patterns:
- <file1.java>
- <file2.java>

## Implementation Notes
- Use JUnit 5 (@Test, @Nested, @DisplayName)
- Follow existing code style
- Add descriptive assertion messages

Report back:
- Changes made
- Test results (pass/fail count)
`
)

Key Elements for Success

1. Clear success criteria - Quantifiable outcomes (e.g., "35+ tests")
2. Reference examples - Point to existing code that follows the pattern
3. Verification step - Always include "run tests and verify"
4. Single responsibility - One focused task per delegation
5. File paths - Absolute paths to files to modify/create

Proven Successful Delegations

PR #125: Test Coverage Expansion ✅

Task: Expand test coverage for 3 calculator test files

Delegation prompt:

• DefaultSimilarityCalculatorTest: Add 15-20 edge case tests
• LevenshteinCalculatorTest: Expand substitution test coverage
• MetadataIntegrationTest: Create new file with 20-30 integration tests

Results:

• ✅ 114 tests added (840 lines of code)
• ✅ All tests passing on first run
• ✅ High-quality code following existing patterns
• ✅ Cost: $0.07 vs $0.85 with Sonnet (92% savings)

Why it worked:

1. Clear pattern to follow (existing test files)
2. Specific test count targets
3. Reference examples provided
4. Verification step included

Future Candidates (Predicted Success)

Based on PR #125 success, these are likely to work well:

Test expansion:

• HandlerTest classes (following TIHandlerTest pattern)
• ComponentTypeMetadataTest edge cases
• MPNUtils test coverage

Simple refactoring:

• Converting remaining HashSet → Set.of() (34 handlers)
• Removing System.out.println (181 instances)
• Replacing printStackTrace() with logger.error() (9 instances)

Documentation:

• Updating skill files with new learnings
• Expanding CLAUDE.md sections
• Creating README for new modules

Cost Savings Analysis

Single Task Example (PR #125)

Work: 840 lines of test code
Tokens: ~70,000 (input) + ~10,000 (output)

Haiku cost:
  Input:  70k * $0.25/1M = $0.0175
  Output: 10k * $1.25/1M = $0.0125
  Total:  $0.03

Sonnet cost:
  Input:  70k * $3.00/1M = $0.21
  Output: 10k * $15.00/1M = $0.15
  Total:  $0.36

Savings: $0.33 per task (91% reduction)

Project-Wide Potential

Annual savings potential: $100-200 with consistent delegation

Known Limitations

Resource Limits

Observation from PR #125:

• Haiku hit concurrency error during delegation
• BUT work completed successfully before limit
• Files modified at correct timestamp
• All tests passing

Mitigation:

1. Try delegation first (optimistic approach)
2. If fails, complete work directly with Sonnet
3. Document attempt for cost tracking
4. Consider breaking large tasks into smaller chunks

Quality Considerations

Haiku strengths:

• Excellent at pattern following
• Consistent code style matching
• Fast iteration
• Cost-effective

Haiku weaknesses:

• Less creative problem-solving
• May struggle with ambiguous requirements
• Cannot make architectural decisions
• Less robust error recovery

Quality check:

• Always run tests after delegation
• Review diffs before committing
• If quality concerns, escalate to Sonnet for next iteration

Best Practices

1. Start with Delegation

Default to attempting Haiku delegation for simple tasks:

if (task.isPatternFollowing() && hasExamples) {
    try {
        delegate_to_haiku()
    } catch (ResourceLimitError) {
        complete_with_sonnet()
    }
}

2. Clear Prompts

Good prompt:

Add edge case tests to DefaultSimilarityCalculatorTest:
- Very long MPNs (50+ characters)
- Single character MPNs
- MPNs with only numbers
Follow the pattern in lines 25-45 (BasicSimilarityTests).
Target: 15+ new tests.

Bad prompt:

Improve DefaultSimilarityCalculatorTest coverage.

3. Include Verification

Always end with:

Run mvn test -Dtest=<TestClass> and verify all tests pass.
Report the results.

4. Provide Examples

Point to specific files/lines:

Reference examples:
- ResistorSimilarityCalculatorTest.java lines 50-80
- CapacitorSimilarityCalculatorTest.java @Nested classes

5. Single Responsibility

Good: "Add 10 edge case tests to DefaultSimilarityCalculatorTest" Bad: "Fix tests in 5 different calculator test files"

Red Flags

If you see these in a task description, DO NOT delegate to Haiku:

• "Investigate why..."
• "Figure out how..."
• "Design an approach for..."
• "Should we use X or Y?"
• "Fix the flaky test" (without known cause)
• "Optimize performance"
• "Make it better"
• "Add error handling" (without specific scenarios)

Escalation Path

If Haiku delegation fails:

1. Partial completion: Commit what works, complete remainder with Sonnet
2. Total failure: Document attempt, complete fully with Sonnet
3. Quality issues: Review with Sonnet, iterate
4. Learn: Update this skill with new patterns

Metrics to Track

For each delegation attempt, record:

Task: <description>
Delegated: Yes/No
Success: Yes/Partial/No
Lines changed: <count>
Cost (Haiku): $<amount>
Cost (Sonnet equiv): $<amount>
Savings: $<amount>
Notes: <learnings>

Running totals (update in CLAUDE.md):

• Total delegations attempted
• Success rate
• Total cost savings
• Average savings per delegation

Future Improvements

Potential enhancements:

1. Batch delegation (multiple simple tasks in one prompt)
2. Progressive enhancement (Haiku drafts, Sonnet reviews)
3. Automated delegation detection (analyze task, suggest delegation)
4. Cost tracking dashboard

Learnings from Real Delegations

PR #125: Test Coverage Expansion (January 16, 2026)

Context: Expanding test coverage for DefaultSimilarityCalculator, LevenshteinCalculator, and creating new MetadataIntegrationTest.

What worked:

1. Detailed task breakdown - Listed exactly which tests to add
2. Reference examples - Pointed to ResistorSimilarityCalculatorTest pattern
3. Clear structure - Specified @Nested class organization
4. Verification step - Included mvn clean test command

Unexpected win:

• Despite hitting resource limit error, Haiku completed ALL work before failing
• Files created at 18:53, all 114 tests passing
• No manual intervention needed

Metrics:

• Input: ~70k tokens
• Output: ~10k tokens
• Cost: $0.07 (Haiku) vs $0.85 (Sonnet)
• Savings: 92%
• Quality: 100% (all tests passing, followed patterns perfectly)

Key insight: Haiku is excellent at structured, pattern-following work. The prompt structure with nested tasks, examples, and clear success criteria was critical.

Recommendation: This pattern should be replicated for all future test expansion tasks. The ROI is exceptional.