4277b10f55
- Neuer Abschnitt "Karten erstellen (Aufgaben-Workflow)" mit Workflow A/B - Bekannte Struktur erweitert: Workflows, Columns, Lanes für Board 1 - Pitfalls ergänzt: Arrival Rule, Parent-Link API, linkedCards read-only - Settings und Plans aktualisiert Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
98 lines
3.7 KiB
Markdown
98 lines
3.7 KiB
Markdown
# Document Model Selection Strategy in CLAUDE.md
|
|
|
|
## Goal
|
|
Create a clear, non-ambiguous rule for model selection (Haiku vs Sonnet vs Opus) that balances token costs with quality, avoiding false economies where cheaper models lead to more expensive debugging cycles.
|
|
|
|
## Key Insight from Discussion
|
|
Using Haiku for execution with Opus for error recovery is likely **more expensive** in practice because:
|
|
- Context loss between models
|
|
- Opus must do blind forensics on Haiku's errors
|
|
- Error recovery costs more tokens than doing it right the first time
|
|
|
|
## Model Characteristics
|
|
|
|
### Haiku (claude-haiku-4-5)
|
|
- **Speed**: Fastest
|
|
- **Cost**: Cheapest per token
|
|
- **Best for**: Truly trivial, zero-risk tasks
|
|
- **Risk**: Higher error rate, weaker context understanding
|
|
|
|
### Sonnet (claude-sonnet-4-5)
|
|
- **Speed**: Medium
|
|
- **Cost**: Medium (~3x Haiku)
|
|
- **Best for**: Most standard tasks, the "default choice"
|
|
- **Risk**: Good balance of speed/quality
|
|
|
|
### Opus (claude-opus-4-5)
|
|
- **Speed**: Slower
|
|
- **Cost**: Highest (~5x Haiku)
|
|
- **Best for**: Complex tasks, deep debugging, critical systems
|
|
- **Risk**: Lowest error rate, best context understanding
|
|
|
|
## Proposed Decision Tree
|
|
|
|
```
|
|
TASK COMPLEXITY:
|
|
|
|
1. Is this trivial? (git status, typo fix, read a single file)
|
|
→ HAIKU
|
|
|
|
2. Is this a standard task with a plan?
|
|
Default: OPUS plans AND executes (one shot, no context loss)
|
|
|
|
Exception: SONNET executes if BOTH:
|
|
- Plan is super-straightforward (mechanical steps, no decisions)
|
|
- AND error probability is very low (documented, tested system)
|
|
|
|
3. Complex/risky tasks?
|
|
→ OPUS (always)
|
|
```
|
|
|
|
**Core Principle:** Avoid model switches mid-task. Context loss costs more than using Opus from start.
|
|
|
|
**Why Opus for execution too?**
|
|
- Real-world execution rarely matches the plan exactly
|
|
- Opus can adapt on-the-fly without re-planning overhead
|
|
- Token-efficient when you factor in error recovery costs
|
|
|
|
## Critical Files
|
|
- `~/.claude/CLAUDE.md` - Global instructions to update
|
|
|
|
## Implementation Plan
|
|
|
|
1. **Add new section to CLAUDE.md** after "Mandatory Workflow"
|
|
- Title: "Model Selection Strategy"
|
|
- Decision tree with concrete examples
|
|
- Anti-patterns (what NOT to do)
|
|
|
|
2. **Document these rules:**
|
|
- **Default: Opus plans AND executes** (one shot, no context loss)
|
|
- **Sonnet**: Only for execution when plan is super-straightforward AND low error risk
|
|
- **Haiku**: Only for truly trivial read-only tasks
|
|
- **Never mix models mid-task** - context loss costs more than using Opus from start
|
|
|
|
3. **Include real examples:**
|
|
- **Haiku**: `git status`, reading a single file, checking if a file exists
|
|
- **Sonnet** (rare exception): Executing a well-tested deployment script when plan is 100% mechanical
|
|
- **Opus** (default): Host scans, documentation, SSH operations, service restarts, debugging, anything with potential for surprises
|
|
|
|
4. **Add anti-pattern:**
|
|
- ❌ "Opus plans, Sonnet/Haiku executes, Opus fixes errors" → Context loss costs more tokens
|
|
- ❌ "Save tokens by using cheaper models" → False economy, error recovery is expensive
|
|
- ✅ "Opus plans AND executes in one shot" → No context loss, adapts on-the-fly
|
|
- ✅ "Only use Haiku for truly trivial read-only tasks"
|
|
|
|
## Verification
|
|
- Read the updated CLAUDE.md section
|
|
- Verify the decision tree is unambiguous
|
|
- Check that examples are concrete, not vague
|
|
- User confirms it's no longer "schwammig"
|
|
|
|
## Decision Made
|
|
After discussion, the strategy is:
|
|
- **Opus as default** for planning AND execution (one shot, no context loss)
|
|
- **Sonnet only as rare exception** when plan is 100% mechanical AND low error risk
|
|
- **Haiku only for trivial read-only** operations
|
|
|
|
Rationale: Context loss between models costs more tokens than using Opus from start. Real-world execution rarely matches plans exactly.
|