---
name: model-router
description: Use when any user request arrives before responding. Classify task complexity and select the right Claude model and effort level before doing anything else.
---

# Model Router

## Overview
Before responding to any request, rate it on two scales: **model** (Haiku / Sonnet / Opus) and **effort** (Low / Medium / High). Then behave accordingly — shallower for simple tasks, deeper for complex ones, and route subagents to cheaper models for grunt work.

## Classification Table

| What the request looks like | Model | Effort |
|-----------------------------|-------|--------|
| Single question, factual lookup, find a file, fix a typo, rename something | **Haiku** | **Low** |
| Add/change a feature, explain a system, fix a moderate bug, write tests, refactor 1–3 files | **Sonnet** | **Medium** |
| Design new architecture, debug across many files, build something from scratch, many unknowns, high stakes | **Opus** | **High** |

### Quick signals — Haiku / Low
- "What does X mean?" / "Where is Y?"
- One-line fix, rename, or typo
- Explain a single concept
- Run a quick command and report back

### Quick signals — Sonnet / Medium
- Change or add a feature to existing code
- Explain how a whole system works
- Standard debugging with a clear reproduction
- Refactoring or writing tests for existing code

### Quick signals — Opus / High
- Architecture or system design from scratch
- Complex bug with unclear root cause spanning many files
- Full feature implementation with many moving parts
- Tasks with high uncertainty or high risk of getting it wrong
- Long multi-step plans or orchestrating many subagents

## Required Steps — Do These Every Time

1. **Announce the rating** at the very top of your first response:
   ```
   > Model: Sonnet · Effort: Medium
   ```

2. **Behave at the rated effort level:**
   - **Low** — Answer directly. One pass. No extra exploration. Skip verification unless obvious.
   - **Medium** — Verify assumptions before acting. Moderate depth. Check your work once at the end.
   - **High** — Full exploration first. Consider edge cases. Try multiple approaches. Thorough verification. Spawn subagents for parallel work.

3. **Assign subagent models to match the subtask, not the overall task:**
   - File reads, quick searches, simple lookups → `model: "haiku"`
   - Code analysis, moderate edits, explanations → `model: "sonnet"`
   - Complex reasoning, architecture decisions → no override (inherits current model)

## Escalation

If you discover mid-task the work is harder than rated, re-announce before continuing:
```
> Escalating to: Opus · High — found cross-system dependency not visible upfront
```

## Model Mismatch Warning

If the task clearly needs Opus but the conversation is running on a lighter model, say so once at the top:
```
> ⚠️ This task is Opus-level. You may get better results restarting in Opus mode.
```
Then do your best with what's available.

## Common Mistakes

| Mistake | Fix |
|---------|-----|
| Rating everything Opus/High | Most requests are Sonnet/Medium. Reserve Opus for genuine complexity. |
| Rating everything Haiku/Low to save tokens | Haiku should not handle multi-file tasks or judgment calls. |
| Spawning Opus subagents for file searches | Always use `model: "haiku"` for mechanical reads and lookups. |
| Forgetting to announce the rating | Always show the `> Model: X · Effort: Y` line first. |
| Not escalating when complexity grows | Re-announce when you discover the task is bigger than it looked. |