Auto-Claude Optimization

Performance tuning, cost reduction, and efficiency improvements.

Performance Overview

Key Metrics

Model Optimization

Model Selection

# Override model in .env
AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929

Extended Thinking Tokens

Configure thinking budget per agent:

# In agent configuration
max_thinking_tokens=5000  # or None to disable

Token Optimization

Reduce Context Size

1. Smaller spec files

   # Keep specs concise
   # Bad: 5000 word spec
   # Good: 500 word spec with clear criteria
   

2. Limit codebase scanning

   # In context/builder.py
   MAX_CONTEXT_FILES = 50  # Reduce from 100
   

3. Use targeted searches

   # Instead of full codebase scan
   # Focus on relevant directories
   

Efficient Prompts

Optimize system prompts in apps/backend/prompts/:

<!-- Bad: Verbose -->
You are an expert software developer who specializes in building
high-quality, production-ready applications. You have extensive
experience with many programming languages and frameworks...

<!-- Good: Concise -->
Expert full-stack developer. Build production-quality code.
Follow existing patterns. Test thoroughly.

Memory Optimization

# Use efficient embedding model
OPENAI_EMBEDDING_MODEL=text-embedding-3-small

# Or offline with smaller model
OLLAMA_EMBEDDING_MODEL=all-minilm
OLLAMA_EMBEDDING_DIM=384

Speed Optimization

Parallel Execution

# Enable more parallel agents (default: 4)
MAX_PARALLEL_AGENTS=8

Reduce QA Iterations

# Limit QA loop iterations
MAX_QA_ITERATIONS=10  # Default: 50

# Skip QA for quick iterations
python run.py --spec 001 --skip-qa

Faster Spec Creation

# Force simple complexity for quick tasks
python spec_runner.py --task "Fix typo" --complexity simple

# Skip research phase
SKIP_RESEARCH_PHASE=true python spec_runner.py --task "..."

API Timeout Tuning

# Reduce timeout for faster failure detection
API_TIMEOUT_MS=120000  # 2 minutes (default: 10 minutes)

Cost Management

Monitor Token Usage

# Enable cost tracking
ENABLE_COST_TRACKING=true

# View usage report
python usage_report.py --spec 001

Cost Reduction Strategies

1. Use cheaper models for simple tasks

   # For simple specs
   AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929 python spec_runner.py --task "..."
   

2. Limit context window

   MAX_CONTEXT_TOKENS=50000  # Reduce from 100000
   

3. Batch similar tasks

   # Create specs together, run together
   python spec_runner.py --task "Add feature A"
   python spec_runner.py --task "Add feature B"
   python run.py --spec 001
   python run.py --spec 002
   

4. Use local models for memory

   # Ollama for memory (free)
   GRAPHITI_LLM_PROVIDER=ollama
   GRAPHITI_EMBEDDER_PROVIDER=ollama
   

Cost Estimation

Memory System Optimization

Embedding Performance

# Faster embeddings
OPENAI_EMBEDDING_MODEL=text-embedding-3-small  # 1536 dim, fast

# Higher quality (slower)
OPENAI_EMBEDDING_MODEL=text-embedding-3-large  # 3072 dim

# Offline (fastest, free)
OLLAMA_EMBEDDING_MODEL=all-minilm
OLLAMA_EMBEDDING_DIM=384

Query Optimization

# Limit search results
memory.search("query", limit=10)  # Instead of 100

# Use semantic caching
ENABLE_MEMORY_CACHE=true

Database Maintenance

# Compact database periodically
python -c "from integrations.graphiti.memory import compact_database; compact_database()"

# Clear old episodes
python query_memory.py --cleanup --older-than 30d

Build Efficiency

Spec Quality = Build Speed

High-quality specs reduce iterations:

# Good spec (fewer iterations)
## Acceptance Criteria
- [ ] User can log in with email/password
- [ ] Invalid credentials show error message
- [ ] Successful login redirects to /dashboard
- [ ] Session persists for 24 hours

# Bad spec (more iterations)
## Acceptance Criteria
- [ ] Login works

Subtask Granularity

Optimal subtask size:

• Too large: Agent gets stuck, needs recovery
• Too small: Overhead per subtask
• Optimal: 30-60 minutes of work each

Parallel Work

Let agents spawn subagents for parallel execution:

Main Coder
├── Subagent 1: Frontend (parallel)
├── Subagent 2: Backend (parallel)
└── Subagent 3: Tests (parallel)

Environment Tuning

Optimal .env Configuration

# Performance-focused configuration
AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929
API_TIMEOUT_MS=180000
MAX_PARALLEL_AGENTS=6

# Memory optimization
GRAPHITI_LLM_PROVIDER=ollama
GRAPHITI_EMBEDDER_PROVIDER=ollama
OLLAMA_LLM_MODEL=llama3.2:3b
OLLAMA_EMBEDDING_MODEL=all-minilm
OLLAMA_EMBEDDING_DIM=384

# Reduce verbosity
DEBUG=false
ENABLE_FANCY_UI=false

Resource Limits

# Limit Python memory
export PYTHONMALLOC=malloc

# Set max file descriptors
ulimit -n 4096

Benchmarking

Measure Build Time

# Time a build
time python run.py --spec 001

# Compare models
time AUTO_BUILD_MODEL=claude-opus-4-5-20251101 python run.py --spec 001
time AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929 python run.py --spec 001

Profile Memory Usage

# Monitor memory
watch -n 1 'ps aux | grep python | head -5'

# Profile script
python -m cProfile -o profile.stats run.py --spec 001
python -c "import pstats; p = pstats.Stats('profile.stats'); p.sort_stats('cumulative').print_stats(20)"

Quick Wins

Immediate Optimizations

1. Switch to Sonnet for most tasks

   AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929
   

2. Use Ollama for memory

   GRAPHITI_LLM_PROVIDER=ollama
   GRAPHITI_EMBEDDER_PROVIDER=ollama
   

3. Skip QA for prototypes

   python run.py --spec 001 --skip-qa
   

4. Force simple complexity for small tasks

   python spec_runner.py --task "..." --complexity simple
   

Medium-Term Improvements

1. Optimize prompts in apps/backend/prompts/
2. Configure project-specific security allowlist
3. Set up memory caching
4. Tune parallel agent count

Long-Term Strategies

1. Self-hosted LLM for memory (Ollama)
2. Caching layer for common operations
3. Incremental context building
4. Project-specific prompt optimization

Related Skills

• auto-claude-memory: Memory configuration
• auto-claude-build: Build process
• auto-claude-troubleshooting: Debugging