--- name: token-efficient-delegation description: Auto-injects token-efficient delegation patterns. Claude tokens are 50x more expensive than MiniMax/GLM/Kimi. --- # Token-Efficient Delegation ## ⛔ CRITICAL: No Claude Subagents **HARD STOP:** Never use `model="haiku"`, `model="sonnet"`, or `model="opus"` in Task tool calls. - Always use: `model` parameter omitted (routes to MiniMax/GLM) - Or explicitly: general-purpose agent (uses cheaper default) - Enforced in: `.claude/settings.local.json` deny rules --- **Core Rule:** Claude = reasoning/decisions, MiniMax/GLM = research/exploration ## The Three Goals (Balance All Three) | Goal | Priority | Strategy | |------|----------|----------| | **Quality** | Keep Opus-level output | Claude synthesizes, subagents gather | | **Cost** | Minimize Claude tokens | Delegate exploration, batch analysis | | **Speed** | Fast execution | Parallel subagents, right model for task | --- ## Quick Decision Tree Before ANY task, ask: 1. **Research/exploration?** → Delegate to MiniMax (fast) or GLM (creative) 2. **Web search needed?** → Delegate to MiniMax 3. **Reading specific known file?** → Claude direct (Read tool) 4. **Analyzing 1-3 images?** → Claude direct 5. **Analyzing 10+ images?** → Parallel subagents (GLM for quality, MiniMax for speed) 6. **Complex reasoning/code?** → Claude direct 7. **Broad codebase search?** → Delegate to general-purpose agent (cheaper default) --- ## Model Characteristics (Official Guidance) | Model | Cost | Speed | Best For | Source | |-------|------|-------|----------|--------| | Claude Opus | 50x | Fast | Complex reasoning, architecture, final decisions | [Anthropic Docs](https://docs.anthropic.com) | | Claude Sonnet | 10x | Fast | Coding, medium reasoning, code review | [Anthropic Docs](https://docs.anthropic.com) | | Claude Haiku | 2x | Fastest | Exploration, classification, simple tasks | [Anthropic Docs](https://docs.anthropic.com) | | MiniMax M2.1 | 1x | Fast | Instruction-following, agents, multi-lang coding | [MiniMax Docs](https://platform.minimax.io) | | GLM-4.7 | 1x | Medium | Creative problem-solving, math, tool use | [Z.AI Docs](https://docs.z.ai) | | GLM-4V | 1x | Medium | Vision, 66K multimodal context, GUI tasks | [Z.AI Docs](https://docs.z.ai) | --- ## When to Use Each Model ### MiniMax M2.1 - Fast Instruction-Follower *Source: [MiniMax Official Docs](https://platform.minimax.io/docs/coding-plan/best-practices)* - Quick web searches - Structured data extraction - File reading with specific questions - Tasks with clear instructions - **8% of Claude cost, 2x speed** - Preserve reasoning chains (return full response with `thinking` field) ### GLM-4.7 - Creative Problem-Solver *Source: [Z.AI GLM-4.7 Docs](https://docs.z.ai/guides/llm/glm-4.7)* - Complex problem exploration - Mathematical reasoning (95.7% on AIME 2025) - Creative solutions needed - Tool use orchestration (42.8%, ties GPT-5.1) - Parallel batch tasks (latency doesn't matter) - Use **thinking mode** for complex tasks, disable for simple queries ### GLM-4V - Vision Tasks *Source: [GLM-V GitHub](https://github.com/zai-org/GLM-V)* - 66K-token multimodal context - Screenshot/GUI analysis - Document parsing (PDFs, charts) - Video understanding with timestamps - Multi-image analysis ### Claude (Haiku/Sonnet/Opus) - Reasoning & Decisions *Source: [Anthropic Subagents Docs](https://code.claude.com/docs/en/sub-agents)* - **Haiku**: Fast exploration, read-only tasks, classification - **Sonnet**: Coding, medium-complexity reasoning - **Opus**: Architecture decisions, synthesis, complex reasoning - Use subagents to isolate verbose output from main context --- ## Quality Preservation Patterns ### Pattern 1: Orchestrator-Worker Split ``` Claude (Orchestrator) → Spawns subagents → Subagents research → Claude synthesizes ``` **Quality preserved because:** Claude makes all decisions based on subagent research. ### Pattern 2: Parallel Research + Single Synthesis ``` 5 MiniMax agents research in parallel → Claude receives summaries → Claude decides ``` **Quality preserved because:** Multiple perspectives, Claude does final reasoning. ### Pattern 3: Two-Stage Review ``` Subagent implements → Reviewer subagent checks → Claude integrates ``` **Quality preserved because:** Cross-validation catches errors before Claude acts. ### Pattern 4: Specific Prompts *Source: [MiniMax Best Practices](https://platform.minimax.io/docs/coding-plan/best-practices)* > "State the 'why' behind your request - when models understand purpose, they provide more accurate answers." ``` BAD: "Research authentication" GOOD: "Find how Godot 4.5 handles input authentication. I need to implement player login. Return: method name, file location, example usage." ``` --- ## Speed Optimization Patterns ### Pattern 1: Parallel Execution Launch independent tasks simultaneously: ``` Task(prompt="Research X") ←─┐ Task(prompt="Research Y") ←─┼─ Single message, parallel execution Task(prompt="Research Z") ←─┘ ``` ### Pattern 2: Right Model for Latency | Need | Model | Latency | |------|-------|---------| | Instant response | Haiku | ~100ms | | Quick research | MiniMax | ~200ms | | Deep analysis | GLM-4.7 | ~500ms | | Complex reasoning | Opus | ~300ms | ### Pattern 3: Batch Over Sequential ``` BAD: Analyze image 1 → wait → Analyze image 2 → wait → ... GOOD: Spawn 10 parallel agents → each analyzes 1 image → aggregate ``` --- ## Cost Optimization Patterns ### Pattern 1: Delegate Exploration *Source: [Anthropic Advanced Tool Use](https://www.anthropic.com/engineering/advanced-tool-use)* > "Delegate verbose operations (tests, log processing, documentation fetching) to subagents so output stays isolated from main conversation." ### Pattern 2: Context Isolation Subagents prevent context bloat: ``` Main context: 5000 tokens (stays lean) Subagent reads 10 files: 15000 tokens (isolated, discarded after) Result passed back: 500 tokens (only what matters) ``` ### Pattern 3: Cascade Routing *Source: [Anthropic Model Selection](https://platform.claude.com/docs/en/docs/about-claude/models/overview)* ``` 60-70% of queries → Haiku (cheapest) 20-30% of queries → Sonnet (medium) 10-15% of queries → Opus (only when needed) ``` ### Pattern 4: Token Suspension via Background Execution **The Problem:** Claude tokens burn while waiting for subagent results. **The Solution:** Use `run_in_background=true` and **end the turn early**. ``` ┌─────────────────────────────────────────────────────────┐ │ BLOCKING (expensive) │ ├─────────────────────────────────────────────────────────┤ │ Claude spawns subagent → waits 60s → gets result │ │ Cost: 60 seconds of Opus tokens BURNED │ ├─────────────────────────────────────────────────────────┤ │ BACKGROUND + END TURN (cheap) │ ├─────────────────────────────────────────────────────────┤ │ Claude spawns with run_in_background=true → ends turn │ │ Cost: ~0 seconds of Opus tokens (turn ended) │ │ Subagent runs: 60s of cheap tokens │ │ User says "continue" → Claude retrieves with TaskOutput│ └─────────────────────────────────────────────────────────┘ ``` **When to suspend (use background):** - Research/exploration tasks >30 seconds - Batch image analysis (10+ images) - Web searches with multiple queries - Code review by subagent **When NOT to suspend:** - Quick lookups (<5 seconds) - Claude needs result to continue current reasoning - Interactive debugging requiring rapid iteration **Fire-and-Retrieve Workflow:** ``` 1. Task(prompt="Research X", run_in_background=true) → Returns immediately with task_id and output_file 2. Claude responds: "Research dispatched. Say 'continue' when ready." 3. User prompts again 4. TaskOutput(task_id="...", block=true) → Returns full results 5. Claude synthesizes and delivers answer ``` **Key phrase:** For tasks >30 seconds, use background execution and end turn early. --- ## Anti-Patterns to Avoid | Anti-Pattern | Problem | Fix | |--------------|---------|-----| | Claude reads 10+ files | Context bloat, expensive | Spawn Explore agent | | Sequential subagent calls | Slow, wasted time | Parallel in single message | | Re-reading subagent results | Duplicates tokens | Trust the summary, decide | | Opus for simple classification | Overkill, expensive | Use Haiku | | GLM for quick lookups | Too slow | Use MiniMax | --- ## Integration with Project Workflows ### HPV (Playtesting) - Use MCP for game state inspection (local, fast) - Delegate screenshot analysis to GLM-4V for batch checks ### Visual Development - 1-3 sprites: Claude direct - 10+ sprites: Parallel GLM-4V agents - Quality gate: Claude reviews subagent findings ### Code Implementation - Research patterns: MiniMax subagents - Write code: Claude direct - Review: Haiku subagent for spec compliance --- ## Official Documentation Sources - [Anthropic Subagents](https://code.claude.com/docs/en/sub-agents) - [Anthropic Advanced Tool Use](https://www.anthropic.com/engineering/advanced-tool-use) - [Anthropic Token-Efficient Tools](https://docs.anthropic.com/en/docs/build-with-claude/tool-use/token-efficient-tool-use) - [MiniMax Best Practices](https://platform.minimax.io/docs/coding-plan/best-practices) - [MiniMax Agent Learnings](https://www.minimax.io/news/minimax-agent-what-we-learned-while-building-in-2025) - [Z.AI GLM-4.7 Docs](https://docs.z.ai/guides/llm/glm-4.7) - [GLM-V Vision Model](https://github.com/zai-org/GLM-V) --- [Claude Opus 4.5 - 2026-01-29]