LangGraph Architect Plugin
Build LangGraph agents faster. Optimize them systematically.
A Claude Code plugin that provides architecture patterns, parallel development workflows, and data-driven optimization for LangGraph applications.
# Install
/plugin marketplace add hiroshi75/langgraph-architect
/plugin install langgraph-architect@langgraph-architect
What You Get
1. Instant LangGraph Expertise
Claude automatically provides architecture guidance when you work with LangGraph:
Build a Gemini+grounding deep-research agent that runs on the CLI using LangGraph.
40+ documentation files covering:
- Core concepts (State, Node, Edge)
- 6 architecture patterns (Routing, Agent, Parallelization, etc.)
- Memory management (Checkpointer, Store, Persistence)
- Tool integration and advanced features
- Model ID references for Claude, Gemini, and OpenAI
2. Parallel Development with Subagents
Break complex graphs into modules. Build them simultaneously.
Your request: "Build a chatbot with intent analysis and RAG search"
Claude decomposes → spawns parallel agents:
├─ langgraph-engineer 1: Intent module (analyze → classify → route)
└─ langgraph-engineer 2: RAG module (retrieve → rerank → generate)
Both run in parallel → integrate into complete graph
3. Prompt-Level Optimization (fine-tune)
The fine-tune skill optimizes your LangGraph prompts without changing graph structure. It activates automatically when Claude detects optimization needs, or invoke manually by /fine-tune.
/fine-tune Fine-Tuning objective: Increase concreteness.
Revise the base prompt so that the generated reports become more concrete and technical, not abstract or generic. Require the model to use specific components, data flows, algorithms, failure modes, and examples.
Use an LLM-based evaluator to assess “concreteness,” and place the evaluation script under eval/. Use that evaluator during tuning.
Auto-activation triggers:
- "improve accuracy", "reduce cost", "optimize prompts"
- "the output is not good enough", "responses are inconsistent"
4-Phase Workflow:
Phase 1: Baseline → Measure current accuracy, latency, cost
Phase 2: Analysis → Identify underperforming nodes and patterns
Phase 3: Optimize → Apply techniques (few-shot, CoT, constraints)
Phase 4: Validate → Statistical validation (3-5 runs) and apply
Typical gains: Accuracy +10-20%, Cost -20-60%
4. Architecture-Level Optimization (arch-tune)
The /arch-tune command explores multiple graph structure improvements in parallel:
/arch-tune "Improve latency to under 2.0s and accuracy to 90%"
What happens:
- Analyze current graph and generate 3-5 improvement proposals
- Implement each proposal in isolated git worktrees (parallel)
- Optimize prompts and parameters for each variant (parallel)
- Compare results with statistical validation
- Merge the winner with your approval
Typical gains: Latency -20-50%, Accuracy +10-30%
Skills
| Skill | Purpose |
|---|
langgraph-architect | Architecture patterns and implementation guidance |
fine-tune | Iterative prompt optimization without changing graph structure |
arch-analysis | Analyze bottlenecks and generate improvement proposals |
Agents
| Agent | Role |
|---|
langgraph-engineer | Implements complete functional modules (2-5 nodes) |
langgraph-tuner | Executes optimization workflow with evaluation |
proposal-comparator | Compares results and recommends best option |
merge-coordinator | Handles user approval and git operations |
Commands
| Command | Description |
|---|
/arch-tune | Full optimization pipeline with parallel exploration |
Quick Examples
Get Architecture Guidance
Just start coding. Claude provides patterns automatically.
# Working on a RAG agent? Claude suggests:
# - retrieve → rerank → generate pattern
# - Checkpointer for conversation memory
# - Subgraph for modular RAG logic
Optimize an Existing Graph
# Prompt-level optimization (no structure changes). It can be auto-triggered or manual by `/fine-tune`.
/fine-tune "Increase accuracy by 15%"
# Architecture-level optimization (structure changes)
/arch-tune "Reduce latency by 30%"
Build Modules in Parallel
For complex applications, Claude spawns multiple langgraph-engineer agents:
