Spaces:
Sleeping
Sleeping
File size: 3,371 Bytes
cc38116 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 | # Model Selection Playbook
> Guidance for choosing models by phase and task type.
>
> **No model is required.** These are recommendations, not requirements.
---
## Selection by Phase
### Planning & Architecture
**Recommended capabilities:**
- Extended reasoning / thinking mode
- Large context window (analyze multiple files)
- Strong at structured output (specs, plans)
**Why:** Planning requires understanding full system context and making architectural decisions.
**Examples:** Models with "thinking" or "reasoning" modes, larger context variants.
---
### Code Implementation
**Recommended capabilities:**
- Fast iteration speed
- Good at code completion
- Tool/function calling (for verification commands)
**Why:** Implementation involves many small changes with frequent verification cycles.
**Examples:** Speed-tier models, code-specialized variants.
---
### Refactoring
**Recommended capabilities:**
- Large context window (see before/after)
- Pattern recognition
- Consistent style application
**Why:** Refactoring requires maintaining consistency across large code changes.
**Examples:** Standard or long-context variants.
---
### Debugging
**Recommended capabilities:**
- Extended reasoning (hypothesis generation)
- Good at reading stack traces
- Context for error patterns
**Why:** Debugging requires hypothesis testing and pattern matching.
**Examples:** Reasoning-focused models.
---
### Code Review
**Recommended capabilities:**
- Large context (review full PR diff)
- Security pattern knowledge
- Style consistency checking
**Why:** Review requires seeing both code and context together.
**Examples:** Long-context variants.
---
## Capability Tiers
| Tier | Characteristics | Best For |
|------|-----------------|----------|
| **Fast** | Quick responses, lower cost | Implementation, iteration |
| **Standard** | Balanced speed/quality | Most tasks |
| **Reasoning** | Extended thinking, slower | Planning, debugging, architecture |
| **Long-context** | >100k tokens | Review, refactoring |
---
## Anti-Patterns
❌ **Using reasoning models for simple edits** — Overkill, slow, expensive
❌ **Using fast models for architecture** — Insufficient depth for complex decisions
❌ **Ignoring context limits** — Leads to quality degradation
❌ **Forcing a specific model** — Breaks model-agnosticism
---
## Model Switching Mid-Session
**When to switch:**
- Context is getting polluted (approaching 50%)
- Task type changes significantly (planning → implementation)
- Current model struggling with task type
**How to switch:**
1. Create state snapshot
2. Update STATE.md with current position
3. Start fresh session with appropriate model
4. Load STATE.md to resume
---
## GSD Model-Agnostic Principle
GSD works with any capable LLM. The methodology compensates for model differences through:
1. **Structured plans** — Reduce ambiguity
2. **Explicit verification** — Catch errors regardless of model
3. **State persistence** — Enable model switching
4. **Fresh context** — Prevent accumulation issues
Choose models based on task needs, not methodology requirements.
---
*See PROJECT_RULES.md for canonical rules.*
*See docs/runbook.md for operational procedures.*
|