Auto-Claude Optimization
Auto-Claude优化
Performance tuning, cost reduction, and efficiency improvements.
| Metric | Impact | Optimization |
|---|
| API latency | Build speed | Model selection, caching |
| Token usage | Cost | Prompt efficiency, context limits |
| Memory queries | Speed | Embedding model, index tuning |
| Build iterations | Time | Spec quality, QA settings |
| 指标 | 影响 | 优化方向 |
|---|
| API latency | 构建速度 | 模型选择、缓存 |
| Token usage | 成本 | 提示词效率、上下文限制 |
| Memory queries | 速度 | 嵌入模型、索引调优 |
| Build iterations | 时间 | 规格质量、QA设置 |
| Model | Speed | Cost | Quality | Use Case |
|---|
| claude-opus-4-5-20251101 | Slow | High | Best | Complex features |
| claude-sonnet-4-5-20250929 | Fast | Medium | Good | Standard features |
| 模型 | 速度 | 成本 | 质量 | 适用场景 |
|---|
| claude-opus-4-5-20251101 | 慢 | 高 | 最佳 | 复杂功能 |
| claude-sonnet-4-5-20250929 | 快 | 中等 | 良好 | 标准功能 |
Override model in .env
Override model in .env
AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929
AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929
Extended Thinking Tokens
扩展思考Token
Configure thinking budget per agent:
| Agent | Default | Recommended |
|---|
| Spec creation | 16000 | Keep default for quality |
| Planning | 5000 | Reduce to 3000 for speed |
| Coding | 0 | Keep disabled |
| QA Review | 10000 | Reduce to 5000 for speed |
为每个Agent配置思考预算：
| Agent | 默认值 | 推荐值 |
|---|
| 规格创建 | 16000 | 保持默认以保证质量 |
| 规划 | 5000 | 降低至3000以提升速度 |
| 编码 | 0 | 保持禁用 |
| QA审核 | 10000 | 降低至5000以提升速度 |
In agent configuration
In agent configuration
max_thinking_tokens=5000 # or None to disable
max_thinking_tokens=5000 # or None to disable
Token Optimization
Token优化
Reduce Context Size
缩小上下文规模
-
Smaller spec files
bash
# Keep specs concise
# Bad: 5000 word spec
# Good: 500 word spec with clear criteria
-
Limit codebase scanning
python
# In context/builder.py
MAX_CONTEXT_FILES = 50 # Reduce from 100
-
Use targeted searches
bash
# Instead of full codebase scan
# Focus on relevant directories
-
更简洁的规格文件
bash
# Keep specs concise
# Bad: 5000 word spec
# Good: 500 word spec with clear criteria
-
限制代码库扫描范围
python
# In context/builder.py
MAX_CONTEXT_FILES = 50 # Reduce from 100
-
使用针对性搜索
bash
# Instead of full codebase scan
# Focus on relevant directories
Optimize system prompts in
:
markdown
<!-- 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.
markdown
<!-- 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.
Use efficient embedding model
Use efficient embedding model
OPENAI_EMBEDDING_MODEL=text-embedding-3-small
OPENAI_EMBEDDING_MODEL=text-embedding-3-small
Or offline with smaller model
Or offline with smaller model
OLLAMA_EMBEDDING_MODEL=all-minilm
OLLAMA_EMBEDDING_DIM=384
OLLAMA_EMBEDDING_MODEL=all-minilm
OLLAMA_EMBEDDING_DIM=384
Enable more parallel agents (default: 4)
Enable more parallel agents (default: 4)
Reduce QA Iterations
减少QA迭代次数
Limit QA loop iterations
Limit QA loop iterations
MAX_QA_ITERATIONS=10 # Default: 50
MAX_QA_ITERATIONS=10 # Default: 50
Skip QA for quick iterations
Skip QA for quick iterations
python run.py --spec 001 --skip-qa
python run.py --spec 001 --skip-qa
Faster Spec Creation
更快的规格创建
Force simple complexity for quick tasks
Force simple complexity for quick tasks
python spec_runner.py --task "Fix typo" --complexity simple
python spec_runner.py --task "Fix typo" --complexity simple
Skip research phase
Skip research phase
SKIP_RESEARCH_PHASE=true python spec_runner.py --task "..."
SKIP_RESEARCH_PHASE=true python spec_runner.py --task "..."
API Timeout Tuning
API超时调优
Reduce timeout for faster failure detection
Reduce timeout for faster failure detection
API_TIMEOUT_MS=120000 # 2 minutes (default: 10 minutes)
API_TIMEOUT_MS=120000 # 2 minutes (default: 10 minutes)
Monitor Token Usage
监控Token使用情况
Enable cost tracking
Enable cost tracking
ENABLE_COST_TRACKING=true
ENABLE_COST_TRACKING=true
View usage report
View usage report
python usage_report.py --spec 001
python usage_report.py --spec 001
Cost Reduction Strategies
成本降低策略
-
Use cheaper models for simple tasks
bash
# For simple specs
AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929 python spec_runner.py --task "..."
-
Limit context window
bash
MAX_CONTEXT_TOKENS=50000 # Reduce from 100000
-
Batch similar tasks
bash
# 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
-
Use local models for memory
bash
# Ollama for memory (free)
GRAPHITI_LLM_PROVIDER=ollama
GRAPHITI_EMBEDDER_PROVIDER=ollama
-
简单任务使用更便宜的模型
bash
# For simple specs
AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929 python spec_runner.py --task "..."
-
限制上下文窗口
bash
MAX_CONTEXT_TOKENS=50000 # Reduce from 100000
-
批量处理相似任务
bash
# 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
-
使用本地模型处理内存
bash
# Ollama for memory (free)
GRAPHITI_LLM_PROVIDER=ollama
GRAPHITI_EMBEDDER_PROVIDER=ollama
| Operation | Estimated Tokens | Cost (Opus) | Cost (Sonnet) |
|---|
| Simple spec | 10k | ~$0.30 | ~$0.06 |
| Standard spec | 50k | ~$1.50 | ~$0.30 |
| Complex spec | 200k | ~$6.00 | ~$1.20 |
| Build (simple) | 50k | ~$1.50 | ~$0.30 |
| Build (standard) | 200k | ~$6.00 | ~$1.20 |
| Build (complex) | 500k | ~$15.00 | ~$3.00 |
| 操作 | 预估Token量 | 成本（Opus） | 成本（Sonnet） |
|---|
| 简单规格 | 10k | ~$0.30 | ~$0.06 |
| 标准规格 | 50k | ~$1.50 | ~$0.30 |
| 复杂规格 | 200k | ~$6.00 | ~$1.20 |
| 构建（简单） | 50k | ~$1.50 | ~$0.30 |
| 构建（标准） | 200k | ~$6.00 | ~$1.20 |
| 构建（复杂） | 500k | ~$15.00 | ~$3.00 |
Memory System Optimization
内存系统优化
Embedding Performance
嵌入性能
Faster embeddings
Faster embeddings
OPENAI_EMBEDDING_MODEL=text-embedding-3-small # 1536 dim, fast
OPENAI_EMBEDDING_MODEL=text-embedding-3-small # 1536 dim, fast
Higher quality (slower)
Higher quality (slower)
OPENAI_EMBEDDING_MODEL=text-embedding-3-large # 3072 dim
OPENAI_EMBEDDING_MODEL=text-embedding-3-large # 3072 dim
Offline (fastest, free)
Offline (fastest, free)
OLLAMA_EMBEDDING_MODEL=all-minilm
OLLAMA_EMBEDDING_DIM=384
OLLAMA_EMBEDDING_MODEL=all-minilm
OLLAMA_EMBEDDING_DIM=384
Limit search results
Limit search results
memory.search("query", limit=10) # Instead of 100
memory.search("query", limit=10) # Instead of 100
Use semantic caching
Use semantic caching
Database Maintenance
数据库维护
Compact database periodically
Compact database periodically
python -c "from integrations.graphiti.memory import compact_database; compact_database()"
python -c "from integrations.graphiti.memory import compact_database; compact_database()"
Clear old episodes
Clear old episodes
python query_memory.py --cleanup --older-than 30d
python query_memory.py --cleanup --older-than 30d
Spec Quality = Build Speed
规格质量 = 构建速度
High-quality specs reduce iterations:
Good spec (fewer iterations)
Good spec (fewer iterations)
Acceptance Criteria
Acceptance Criteria
Bad spec (more iterations)
Bad spec (more iterations)
Acceptance Criteria
Acceptance Criteria
Optimal subtask size:
- Too large: Agent gets stuck, needs recovery
- Too small: Overhead per subtask
- Optimal: 30-60 minutes of work each
最优子任务规模：
- 过大：Agent陷入停滞，需要恢复
- 过小：子任务 overhead 过高
- 最优：每个子任务对应30-60分钟工作量
Let agents spawn subagents for parallel execution:
Main Coder
├── Subagent 1: Frontend (parallel)
├── Subagent 2: Backend (parallel)
└── Subagent 3: Tests (parallel)
让Agent生成子Agent以并行执行：
Main Coder
├── Subagent 1: Frontend (parallel)
├── Subagent 2: Backend (parallel)
└── Subagent 3: Tests (parallel)
Optimal .env Configuration
最优.env配置
Performance-focused configuration
Performance-focused configuration
AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929
API_TIMEOUT_MS=180000
MAX_PARALLEL_AGENTS=6
AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929
API_TIMEOUT_MS=180000
MAX_PARALLEL_AGENTS=6
Memory optimization
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
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
Reduce verbosity
DEBUG=false
ENABLE_FANCY_UI=false
DEBUG=false
ENABLE_FANCY_UI=false
Limit Python memory
Limit Python memory
export PYTHONMALLOC=malloc
export PYTHONMALLOC=malloc
Set max file descriptors
Set max file descriptors
time python run.py --spec 001
time python run.py --spec 001
Compare models
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
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
分析内存使用
Monitor memory
Monitor memory
watch -n 1 'ps aux | grep python | head -5'
watch -n 1 'ps aux | grep python | head -5'
Profile script
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)"
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)"
Immediate Optimizations
即时优化措施
-
Switch to Sonnet for most tasks
bash
AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929
-
Use Ollama for memory
bash
GRAPHITI_LLM_PROVIDER=ollama
GRAPHITI_EMBEDDER_PROVIDER=ollama
-
Skip QA for prototypes
bash
python run.py --spec 001 --skip-qa
-
Force simple complexity for small tasks
bash
python spec_runner.py --task "..." --complexity simple
-
大部分任务切换为Sonnet模型
bash
AUTO_BUILD_MODEL=claude-sonnet-4-5-20250929
-
使用Ollama处理内存
bash
GRAPHITI_LLM_PROVIDER=ollama
GRAPHITI_EMBEDDER_PROVIDER=ollama
-
原型开发跳过QA
bash
python run.py --spec 001 --skip-qa
-
小型任务强制设置为简单复杂度
bash
python spec_runner.py --task "..." --complexity simple
Medium-Term Improvements
中期改进措施
- Optimize prompts in
- Configure project-specific security allowlist
- Set up memory caching
- Tune parallel agent count
- 优化中的提示词
- 配置项目专属安全白名单
- 启用内存缓存
- 调整并行Agent数量
- Self-hosted LLM for memory (Ollama)
- Caching layer for common operations
- Incremental context building
- Project-specific prompt optimization
- 自托管LLM处理内存（Ollama）
- 为常见操作添加缓存层
- 增量式上下文构建
- 项目专属提示词优化
- auto-claude-memory: Memory configuration
- auto-claude-build: Build process
- auto-claude-troubleshooting: Debugging
- auto-claude-memory: 内存配置
- auto-claude-build: 构建流程
- auto-claude-troubleshooting: 调试