model-selection

name	model-selection
version	1.0.0
category	ai
description	Guide AI model selection based on task complexity, cost constraints, and latency requirements
type	reference
capabilities	[]
requires	[]
see_also	[]
tags	[]
scripts_exempt	true

Model Selection Skill

Version: 1.0.0 Category: Optimization Triggers: Starting tasks, choosing Codex model, usage optimization

Quick Reference

Model Selection Decision Tree

NEW TASK
    │
    ├── WORK REPO + COMPLEX → OPUS
    ├── WORK REPO + STANDARD → SONNET
    ├── PERSONAL + SIMPLE → HAIKU
    └── DEFAULT → SONNET

Quick Selection Guide

Model	Target %	Use For
OPUS	30%	Architecture, multi-file refactoring (>5 files), security review
SONNET	40%	Standard implementations, code review, documentation
HAIKU	30%	Quick queries, status checks, simple operations

Automated Model Suggestion

# Get model recommendation before each task
./scripts/monitoring/suggest_model.sh <repository> "<task description>"

# Examples:
./scripts/monitoring/suggest_model.sh digitalmodel "Design authentication architecture"
# → Recommends: OPUS (complexity score: 4)

./scripts/monitoring/suggest_model.sh digitalmodel "Implement user login"
# → Recommends: SONNET (complexity score: 1)

./scripts/monitoring/suggest_model.sh hobbies "Quick file check"
# → Recommends: HAIKU (complexity score: -3)

Complexity Scoring

Algorithm evaluates:

1. Keywords - architecture/refactor → +3, implement/feature → +1, check/status → -2
2. Repository Tier - Work Tier 1 → +1, Personal → -1
3. Task Length - >15 words → +1, <5 words → -1

Score Mapping:

• Score ≥3: OPUS
• Score 0-2: SONNET
• Score <0: HAIKU

Repository Tiers

Work Repositories

Tier 1 (Production): 60% Opus, 30% Sonnet, 10% Haiku

• digitalmodel, energy, client-a

Tier 2 (Active): 30% Opus, 50% Sonnet, 20% Haiku

• assetutilities, worldenergydata

Tier 3 (Maintenance): 10% Opus, 30% Sonnet, 60% Haiku

• lng-a, client-d, OGManufacturing

Personal Repositories

Active: 20% Opus, 40% Sonnet, 40% Haiku Experimental: 5% Opus, 25% Sonnet, 70% Haiku Archive: 0% Opus, 20% Sonnet, 80% Haiku

Usage Monitoring

Check before starting work: https://Codex.ai/settings/usage

Alert Thresholds:

• Sonnet >70% → Switch to Opus/Haiku
• Session >80% → Batch work or wait
• Overall >80% → Defer non-critical

OPUS Use Cases

✅ Multi-file refactoring (>5 files) ✅ Architecture decisions ✅ Complex algorithm design ✅ Security-critical code review ✅ Cross-repository coordination ✅ Performance optimization strategies

SONNET Use Cases

✅ Standard feature implementation ✅ Code review (single PR) ✅ Documentation writing ✅ Test generation ✅ Bug fixing (standard complexity) ✅ Configuration updates

HAIKU Use Cases

✅ File existence checks ✅ Simple grep/search operations ✅ Quick status updates ✅ Log analysis (pattern matching) ✅ Template generation ✅ Format validation

Emergency Protocols

If Sonnet >80%

⛔ STOP using Sonnet immediately
✅ Switch to Opus for critical work
✅ Switch to Haiku for everything else
📅 Defer non-urgent work to Tuesday

If Session >80%

⏸️  Pause AI tasks
⏰ Wait for session reset (~3-4 hours)
📦 Batch work for next session

Full Reference

See: @docs/AI_MODEL_SELECTION_AUTOMATION.md See: @docs/CLAUDE_MODEL_SELECTION_QUICK_REFERENCE.md

---

Use this when starting tasks, selecting models, or optimizing AI usage.