Model Selection

ROUTINE TASKS → Sonnet 4.6
- Implement a feature following an existing pattern
- Write tests for existing code
- Fix a bug with a clear stack trace
- Add a new endpoint matching the project's convention
- Run scaffolding or code generation
- Apply a known refactoring pattern
- Format, lint, or fix build errors
- Write documentation from existing code

COMPLEX TASKS → Opus 4.6
- Design a new module or subsystem from scratch
- Choose between architecture approaches (VSA vs CA vs DDD)
- Debug a subtle issue with no clear stack trace
- Refactor with trade-offs (performance vs readability, consistency vs simplicity)
- Review architecture for design flaws
- Make decisions with incomplete or conflicting requirements
- Untangle complex dependencies or circular references
- Write a migration strategy for breaking changes

SIMPLE TASKS → Haiku 4.5 (subagents only)
- Look up a file path or symbol location
- Run a build/test and report results
- Search for a pattern across files
- Summarize a file or module
- Format or validate data

PHASE 1: PLAN (Opus 4.6)
├── Analyze requirements and constraints
├── Identify architectural trade-offs
├── Design the approach with rationale
├── Define acceptance criteria
└── Output: detailed implementation plan

PHASE 2: EXECUTE (Sonnet 4.6)
├── Implement following the Opus plan
├── Write code, tests, configurations
├── Run build and test verification
├── Handle routine issues (compilation errors, test fixes)
└── Output: working implementation

PHASE 3: REVIEW (Opus 4.6)
├── Review implementation against the plan
├── Check for subtle issues (race conditions, N+1, security)
├── Evaluate architectural compliance
├── Suggest refinements
└── Output: approval or specific revision requests

Claude Code CLI:
- /model opus    → Switch to Opus 4.6 (planning/review phases)
- /model sonnet  → Switch to Sonnet 4.6 (implementation phase)
- /model auto    → Let Claude Code choose based on task

Toggle fast mode:
- /fast          → Toggle fast mode (Opus fast output for throughput)

SUBAGENT: "Find all authentication middleware in the project"
→ MODEL: Haiku 4.5
→ WHY: Simple search task, no reasoning required

SUBAGENT: "Run dotnet test and summarize failures"
→ MODEL: Haiku 4.5
→ WHY: Execute command, parse output, no complex analysis

SUBAGENT: "Analyze the dependency graph for circular references and suggest fixes"
→ MODEL: Sonnet 4.6
→ WHY: Needs to understand project structure and propose solutions

SUBAGENT: "Review this PR for architectural issues and security vulnerabilities"
→ MODEL: Opus 4.6
→ WHY: Deep reasoning about trade-offs, subtle issue detection

SUBAGENT: "Summarize what the Orders module does"
→ MODEL: Haiku 4.5 or Sonnet 4.6
→ WHY: Haiku for a quick overview, Sonnet if the module is complex

// BAD — Opus for a routine CRUD endpoint
Using Opus 4.6 to implement GetOrderById following the exact same pattern
as the existing GetCustomerById. No decisions to make, just pattern replication.
*Slower and more expensive than needed*

// GOOD — Sonnet for pattern replication
Using Sonnet 4.6 to implement GetOrderById. The pattern is established,
the code is straightforward, and Sonnet executes it faster.

// BAD — Sonnet for "should we use VSA or Clean Architecture?"
Sonnet gives a reasonable answer but may miss nuanced trade-offs
about team size, domain complexity, and long-term maintenance.
*The wrong architecture costs months of refactoring*

// GOOD — Opus for architectural decisions
Opus weighs team size, domain complexity, current codebase patterns,
and long-term implications. The architecture decision is worth the
extra reasoning power.