model-selector

Model Selector

Evaluate the incoming task and recommend the optimal Claude model and effort level before proceeding. State the recommendation clearly, explain the reasoning in one sentence, and provide the switch command.

---

Model Tiers

Model	Alias	Best For
Haiku	haiku	Simple edits, grep, formatting, docs, single-file changes
Sonnet	sonnet	Standard coding, debugging, moderate refactors, most daily work
Opus	opus	Complex architecture, deep multi-file reasoning, ambiguous specs, long agentic runs

Effort Levels

Level	Flag	When
Low	--effort low	Routine tasks, clear requirements, simple edits
Medium	--effort medium	Standard feature work, debugging with stack trace
High	--effort high	Non-obvious design decisions, trade-off analysis
Max	--effort max	Architecting systems, deep bugs, security review

Note: effort is supported on Opus 4.6 and Sonnet 4.6 only.

---

Classification Heuristics

→ Haiku if the task is:

• Single file, localised change (rename, reformat, add comment)
• Purely mechanical (find/replace, generate boilerplate from template)
• Documentation only
• Answering a narrow factual question about the codebase

→ Sonnet if the task is:

• Multi-file but well-scoped feature work
• Debugging with a clear error and stack trace
• Writing or updating tests
• Standard API integration or CRUD work
• Most day-to-day development tasks

→ Opus if the task is:

• Ambiguous or under-specified requirements needing refinement
• Architectural decisions spanning multiple services/systems
• Long agentic runs (>10 tool calls expected)
• Security-sensitive review or threat modelling
• Performance profiling with non-obvious root cause
• Migrating or refactoring large, complex codebases

Elevate effort level if:

• Multiple valid design approaches exist and trade-offs are non-trivial
• The task involves concurrency, distributed state, or security
• Previous attempts at this task failed unexpectedly

---

Output Format

## Model Recommendation

**Model**: <Model Name> (`/model <alias>`)
**Effort**: <Low | Medium | High | Max> (`--effort <level>` or "default is fine")
**Reason**: <One sentence justification>

**To apply**:
  /model <alias>

If already on the correct model and effort, confirm and proceed immediately.

---

Behaviour Notes

• Do not ask the user to classify their own task — infer from what they've said
• If task scope is genuinely unclear, ask one clarifying question before recommending
• Default to Sonnet when uncertain — it covers the widest range well
• Never recommend Opus for tasks that are clearly Haiku or Sonnet scope
• Models can be switched mid-session with /model <alias> — no restart needed
• After giving the recommendation, stop and wait for the user to respond before proceeding — do not continue with the task until the user confirms or replies