Week 35.5 | Raw with Claude

// The original validation (too strict)
private fun isValidPhoneNumber(phone: String): Boolean {
    return phone.matches(Regex("^\\+1\\d{10}$"))
    // Requires: +1 followed by exactly 10 digits
}

// After multiple iterations
private fun isValidPhoneNumber(phone: String): Boolean {
    val normalized = normalizePhoneNumber(phone)
    return normalized.length in 10..11 && 
           normalized.all { it.isDigit() }
}

private fun normalizePhoneNumber(phone: String): String {
    return phone.filter { it.isDigit() }
}

# .claude/subagents/shokan_ui_claude.md
You are the UI implementation specialist for Shokken app.
Focus on Compose Multiplatform UI components and Material 3 design.
Never modify business logic or data layer code.

# .claude/subagents/shokan_validator_claude.md
You run tests and fix violations.
You are responsible for Konsist, Detekt, and KTLint compliance.
Always run the full test suite before proposing fixes.

# .claude/subagents/shokan_researcher_claude.md
You research best practices and library documentation.
Always provide multiple options with pros/cons analysis.
Ground all recommendations in actual documentation.

// Single investigation by subagent
data class TokenUsage(
    val operation: String,
    val tokensUsed: Int
)

val actualUsage = listOf(
    TokenUsage("Initial investigation", 15_000),
    TokenUsage("Research best practices", 20_000),
    TokenUsage("Implementation attempt", 8_000),
    TokenUsage("Running tests", 48_000), // Nearly half the context!
    TokenUsage("Fixing test failures", 12_000)
)
// Total: 103,000 tokens (80% of Opus limit)

# Problems encountered:
1. No local emulator support:
   - Emulator runs on server, not visible locally
   - Must use physical device for testing
   
2. Terminal flashing issues:
   - Requires complete tab restart (not just Claude restart)
   
3. IDE cache invalidation danger:
   - "Invalidate Cache and Restart" bricks the installation
   - Remote development still in beta
   
4. Build/test delays:
   - Commands execute on remote server
   - Network latency adds overhead

# Iteration 1: Misdiagnosis
"Add normalization before validation" 
# Result: Normalized for backend, didn't fix UI issue

# Iteration 2: Wrong scope
"Research included international phone support"
# Result: Suggested platform-specific libraries (rejected)

# Iteration 3: Constrained research
"US-only, no new libraries, presentation layer only"
# Result: Simple validation fix

# Iteration 4: Test failures
"Konsist detecting hard-coded strings"
# Multiple rounds of fixes creating new violations

// Pre-commit hooks (3+ minutes to run)
tasks.register("preCommitHooks") {
    dependsOn(
        "ktlintCheck",        // Code style
        "detekt",             // Static analysis
        "konsistTest",        // Architecture rules
        "test",               // Unit tests
        "koverVerify"         // Coverage requirements
    )
}

// Konsist rules catching issues
Konsist.scopeFromProject()
    .classes()
    .assert { clazz ->
        // No hard-coded strings (even numbers!)
        !clazz.text.contains(Regex("\"\\d+\""))
    }
    
// The frustration: Konsist includes comments!
// "123456789" in a comment fails the test

sealed class ClaudeQuirk {
    object EnthusiasmTrap : ClaudeQuirk() {
        // "I found the issue!" - 30-40% false positive rate
    }
    
    object WrongAgentSelection : ClaudeQuirk() {
        // UI agent tries to fix test failures
        // Validator agent attempts implementations
    }
    
    object ContextAmnesia : ClaudeQuirk() {
        // Forgets instructions from earlier in conversation
        // Requires periodic reassertion of rules
    }
    
    object OvereagerImplementation : ClaudeQuirk() {
        // Implements before investigating
        // Requires explicit "investigation only" instructions
    }
}

subscription_details:
  cost: $200/month
  model: Claude Opus (200K context)
  limits:
    - Not unlimited usage
    - Rate limits reset every 5 hours (currently)
    - After Sept 1: Weekly reset (concerning!)
    - Multiple windows drain quota quickly
  
  actual_productivity:
    - Can hit limits in 1-2 hours with heavy usage
    - 6-8 concurrent windows: probably 1 hour max
    - Need to manage token usage strategically

github_actions:
  free_tier: 2000 minutes/month
  
  minute_costs:
    ubuntu_runner: 1x multiplier (1 minute = 1 minute)
    macos_runner: 10x multiplier (1 minute = 10 minutes!)
  
  ios_build_problem:
    - Need macOS runner for iOS builds
    - 200 minutes actual = 2000 minutes billed
    - Can't test iOS regularly without paying
    
  current_status: "Out of minutes, all CI/CD failing"

enum class DeveloperRole {
    BEFORE_AI {
        override val responsibilities = listOf(
            "Syntax generator",
            "Documentation reader",
            "Implementation details focus"
        )
    },
    
    WITH_AI {
        override val responsibilities = listOf(
            "Product manager mindset",
            "Architecture decisions",
            "Quality gatekeeper",
            "AI shepherd"
        )
    }
}

// The skill transfer
data class SkillEvolution(
    val declining: List<String> = listOf(
        "Memorizing syntax",
        "Reading regex",
        "Manual refactoring"
    ),
    val emerging: List<String> = listOf(
        "Prompt engineering",
        "Context management",
        "Multi-agent orchestration",
        "Iterative problem refinement"
    )
)