← Discover MCPs and Agents
c
AgentAI & MLGitHub

claudecode_steering

Comprehensive steering framework for Claude Code with GPT-5.2 MCP agents - capture your development standards once, enforce them forever

Links

README

From the repo.

Claude Code Steering Framework

Guidance that persists. Standards that enforce.

What is This?

This project is a comprehensive steering framework for Claude Code that preserves your development style, standards, and guidance in documentation. Instead of repeatedly instructing AI assistants about your preferences, this framework:

  • Captures your guidance in structured documentation
  • Enforces your standards automatically through MCP agents
  • Preserves your style across all development sessions
  • Scales your expertise via GPT-5.2-powered consultants

Core Philosophy: Guidance that persists. Standards that enforce.

What is Steering?

Steering is the practice of capturing your development guidance in version-controlled documentation, then having AI agents automatically enforce it during every code change.

Traditional approach:

  • Manually review every pull request
  • Repeatedly explain the same standards in code reviews
  • Hope developers remember your preferences
  • Standards drift over time as context is lost

Steering approach:

  • Document your standards once in docs/
  • GPT-5.2 agents read and enforce them automatically
  • Standards evolve with git history
  • New patterns discovered → added to docs → enforced everywhere

Key difference: Your expertise becomes infrastructure, not tribal knowledge.

Origins

This framework emerged from a real Python project (0 to ~200,000 lines of code) where development principles were captured and refined through iterative steering with Claude Code. The project specifics have been stripped away, but the fundamental software development principles remain:

  • Language-agnostic principles - Testability, type safety, clear abstractions, systematic debugging
  • Python-specific practices - Modern type hints, Pydantic schemas, ABC patterns, inheritance-based fakes
  • Battle-tested patterns - Every rule in docs/ solved a real problem or prevented a real bug

The documentation is structured to separate universal principles from language-specific rules, making it easy to adapt this framework to other languages (TypeScript, Go, Rust, etc.) while preserving the core development philosophy.

Quickstart (5 Minutes)

This repository is a template/steering pack for Claude Code. Here's what you get:

  1. Pre-configured MCP agents - Three GPT-5.2 specialist agents ready to enforce your standards
  2. Comprehensive documentation - 5 detailed guides covering code style, testing, architecture, troubleshooting, and infrastructure
  3. Workflow automation - Claude Code automatically consults agents at key development steps
  4. Self-improving - Documentation updates with every new pattern discovered

What's runnable today:

  • Read the documentation in docs/ to understand the coding standards
  • Review MCP agent configurations in prompts/
  • Start a Claude session and ask it to plan a feature - it will automatically consult the agents
  • Clone this as your project template and customize the standards to your preferences

What you'll add:

  • src/ - Your application code (following docs/code-style.md)
  • tests/ - Your test suite (following docs/test-style.md)

Adapting to other languages:

  • Documentation is split into Fundamental Principles (language-agnostic) and Python-Specific Rules
  • Keep the fundamental principles sections intact
  • Ask Claude Code: "Translate the Python-specific sections to TypeScript/Go/Rust"
  • Review Claude's translation and enhance as needed
  • Update MCP agent prompts to reference your language standards

Example workflow for TypeScript:

User: "Read docs/code-style.md and translate all Python-specific sections to TypeScript. Keep the fundamental principles unchanged."

Claude: [Produces TypeScript-specific type system rules, testing patterns, etc.]

User: "Review and enhance the TypeScript translation with best practices."

Claude: [Refines the translation, adds TypeScript idioms]

How It Works

1. Documentation as Steering Mechanism

All your development preferences are documented in:

  • docs/code-style.md - Your coding standards (type safety, patterns, architecture)
  • docs/test-style.md - Your testing philosophy (no mocks, inheritance-based fakes, 90% coverage)
  • docs/troubleshooting.md - Your debugging approaches and common solutions
  • docs/architecture.md - Your system architecture and design patterns
  • docs/infrastructure-style.md - Your infrastructure standards (Terraform/IaC, CI/CD, deployment)
  • CLAUDE.md - Your workflow and process guidance for Claude Code

2. Automated Enforcement via MCP Agents

Three GPT-5.2 specialist agents automatically enforce your standards:

AgentMCP Server NameWhenPurpose
GPT-Architectgpt5-architectStep 0, Step T-1Validates design decisions against your architectural principles
GPT-Reviewergpt5-reviewerStep T, T+1, L-1Ensures code compliance with your standards
GPT-Troubleshootergpt5-troubleshooterWhen debuggingApplies your systematic troubleshooting methodology

Note: Display names (GPT-Architect) and MCP server names (gpt5-architect) refer to the same agents.

These agents read your documentation and enforce it on every feature, refactoring, or bug fix.

3. Self-Improving Documentation

Every development session updates the documentation (Step L):

  • New patterns discovered → Added to code-style.md
  • Bugs encountered → Added to troubleshooting.md
  • Architectural decisions → Captured in architecture.md

Result: Your guidance gets stronger over time, never repeats, always available.

Prerequisites

Required Tools

  1. Claude Code - The AI-powered CLI for software engineering

  2. Codex CLI - MCP server runtime for GPT-5.2 agents

  3. Node.js - For Context7 MCP agent

    • Required to run npx commands
    • Version: 18+ recommended
  4. Python - 3.12+

    • For project code (when you start writing actual application code)
  5. Git - For version control and worktree workflow

  6. Context7 API Key - For library documentation lookup

Installation

Step 1: Install Claude Code

# macOS (via Homebrew)
brew install --cask claude-code

# Or download from: https://claude.com/code

Verify installation:

claude --version

Step 2: Install Codex CLI

# macOS (via Homebrew)
brew install --cask codex

# Or download from: https://github.com/openai/codex

Verify installation:

codex --version  # Should show v0.72.0+

Configure Codex with your OpenAI API key (for GPT-5.2 access):

# Set up authentication
codex login

# Or set environment variable
export OPENAI_API_KEY="your-api-key"

Step 3: Clone This Repository

git clone https://github.com/your-username/claudecode-steering-blog.git
cd claudecode-steering-blog

Note: Replace your-username with your actual GitHub username if you've forked this repo.

Step 4: Configure Context7 API Key

Context7 provides library documentation lookup while coding

1. Get API Key: Visit https://context7.ai and sign up for a free API key

2. Set environment variable:

# Add to ~/.bashrc, ~/.zshrc, or equivalent
export CONTEXT7_API_KEY="your-api-key-here"

# Reload shell
source ~/.bashrc  # or ~/.zshrc

3. Verify:

echo $CONTEXT7_API_KEY  # Should display your key

Context7 is already configured in .mcp.json and will be available once the environment variable is set.

Step 5: Verify MCP Agents

Test that all MCP agents are configured:

# From project root
claude mcp list

You should see:

  • ✅ gpt5-architect - Connected
  • ✅ gpt5-reviewer - Connected
  • ✅ gpt5-troubleshooter - Connected

Note: claude mcp list only checks connectivity. To verify authentication and full functionality, start a Claude session and the agents will be tested when first invoked.

Project Structure

claudecode-steering-blog/
├── docs/                      # Your steering documentation
│   ├── code-style.md         # Coding standards
│   ├── test-style.md         # Testing standards
│   ├── troubleshooting.md    # Debug procedures
│   ├── architecture.md       # System architecture
│   └── infrastructure-style.md # Infrastructure standards (Terraform, CI/CD)
├── prompts/                   # MCP agent system prompts
│   ├── architect.txt         # Architect agent configuration
│   ├── reviewer.txt          # Reviewer agent configuration
│   └── troubleshooter.txt    # Troubleshooter agent configuration
├── scripts/                   # MCP agent launchers
│   └── gpt5-agent.sh         # Unified script for all three GPT-5 agents
├── .mcp.json                  # MCP server configuration
├── CLAUDE.md                  # Claude Code workflow guidance
├── AGENTS.md                  # Agent usage notes
└── README.md                  # This file

Note: src/ and tests/ directories will be created when you start writing code.

Usage

Start a Development Session

# Use git worktrees for feature isolation (recommended)
git worktree add ../myproject-feature -b feature-name
cd ../myproject-feature

# Start Claude Code
claude

Claude will:

  1. Read CLAUDE.md for workflow guidance
  2. Consult GPT-Architect when planning features
  3. Consult GPT-Reviewer before/after tests
  4. Enforce your standards automatically
  5. Update documentation with new patterns (Step L)

Git Worktree Workflow

ALWAYS use git worktrees for feature development - this is enforced in CLAUDE.md

Why worktrees:

  • Isolate each feature in its own directory
  • Switch between features instantly (no stashing)
  • Work on multiple features in parallel
  • Keep main branch clean

Setup a new feature:

# From your main project directory
cd /path/to/claudecode-steering-blog

# Create worktree for new feature
git worktree add ../claudecode-steering-blog-feature-name -b feature-name

# Move into the worktree
cd ../claudecode-steering-blog-feature-name

# Start Claude Code
claude

List your worktrees:

git worktree list

Remove a worktree when done:

git worktree remove ../claudecode-steering-blog-feature-name

Planning Mode: Always Plan Before Editing

CRITICAL: Use planning mode for all non-trivial changes

When Claude suggests changes, ALWAYS use planning mode first:

  1. Trigger planning mode - Claude will enter plan mode for complex tasks
  2. Review the plan - Inspect proposed changes before any code is written
  3. Provide feedback - Ask questions, request changes to approach
  4. Approve or reject - Only accept plan when satisfied
  5. Execution - Claude implements after plan approval

Benefits:

  • See the approach before code is written
  • Catch design issues early
  • No wasted effort on wrong direction
  • Clear understanding of changes

In Claude Code:

  • Planning mode is triggered automatically for complex tasks
  • You can always request to see a plan first
  • Plans include file changes, new files, testing approach
  • MCP agents (Architect, Reviewer) are consulted during planning

Context7: Library Documentation Lookup

Context7 provides instant access to library documentation while coding

Note: Context7 API key should already be configured from Step 4 of Installation. If not, see that section.

Using Context7

During development, Claude can:

  • Look up library documentation automatically
  • Get latest API references
  • Find usage examples
  • Check compatibility

Example:

User: "Use httpx for API calls"

Claude (via Context7):
  ├─ Looks up httpx documentation
  ├─ Finds async client examples
  ├─ Checks latest version compatibility
  └─ Implements with best practices

Manual lookup:

User: "What's the context7 documentation for Pydantic?"
Claude: [Uses context7 MCP agent to fetch Pydantic docs]

Supported libraries: Most popular Python, JavaScript, and other languages

Typical Development Flow

User: "Add user authentication feature"

Claude:
  ├─ Reads docs/code-style.md, docs/architecture.md
  ├─ Creates implementation plan
  ├─ Step 0: Consults GPT-Architect (validates design)
  ├─ Implements feature following your standards
  │   ├─ NewType for UserIds
  │   ├─ Pydantic schemas for data
  │   ├─ ABC (not Protocol) for interfaces
  │   ├─ Dependencies passed as parameters (testable)
  │   └─ Context managers for resources
  ├─ Step T: Consults GPT-Reviewer (pre-test review)
  ├─ Writes tests (inheritance-based fakes, no mocks)
  ├─ Step T+1: Consults GPT-Reviewer (post-test review)
  ├─ Step L-1: Final review with GPT-Reviewer
  └─ Step L: Updates documentation with new patterns

MCP Agents in Action

GPT-Architect (Step 0, T-1):

Reviewing your plan for user authentication...

✅ Approvals:
- Business logic testable (dependencies passed as parameters)
- NewType(UserId) defined in src/utils/types.py
- Pydantic schemas for User, Credentials
- Using ABC, not Protocol ✓

⚠️ Concerns:
- src/auth/service.py:45 - Thin wrapper detected
  Recommendation: Expose auth.sessions.create() directly

💡 Recommendations:
- Consider ABC for shared session logic
- Check if auth library already exists (httpx-auth, authlib)?

GPT-Reviewer (Step T, T+1, L-1):

Reviewing code changes...

✅ Approvals:
- Type safety: All IDs use NewType ✓
- Modern syntax: Using list[str] not List[str] ✓

🐛 Bugs Found:
- src/auth/service.py:67 - Unhandled None return
  Fix: Add None check before accessing result.user_id

⚠️ Standards Violations:
- tests/unit/auth/test_service.py:12 - Using unittest.mock
  Fix: Replace with inheritance-based FakeAuthService

GPT-Troubleshooter (When debugging):

🔍 Hypothesis:
Database connection not properly closed in error path

✅ Verification Plan:
1. Add logging at auth/service.py:45 (before DB call)
2. Add logging at auth/service.py:52 (after DB call)
3. Run: grep -r "Database(" src/ (find similar patterns)

⚡ Speed Up Debugging:
pytest -k "test_auth_failure" -v

Documentation Standards

Your Guidance is Captured In:

docs/code-style.md - Enforces:

  • Zero Any types in business logic
  • NewType for all domain IDs
  • Pydantic schemas for all data contracts
  • Orchestrator pattern for testability
  • ABC for >50% code reuse
  • No Protocol in internal code
  • No thin wrappers

docs/test-style.md - Enforces:

  • No unittest.mock or monkeypatch
  • Inheritance-based fakes only
  • 90% coverage target on business logic
  • Bug fixes must include test cases

docs/troubleshooting.md - Guides:

  • 8 systematic troubleshooting principles
  • Search for similar bugs before fixing
  • Minimize time to reproduce
  • Common error patterns with solutions

Customizing for Your Project

Update Your Standards

Edit the documentation files to match your preferences:

# Edit coding standards
vim docs/code-style.md

# Edit testing standards
vim docs/test-style.md

# Edit troubleshooting approaches
vim docs/troubleshooting.md

The MCP agents will automatically enforce your updated standards.

Customize MCP Agents

Edit the agent system prompts:

# Customize architect behavior
vim prompts/architect.txt

# Customize reviewer behavior
vim prompts/reviewer.txt

# Customize troubleshooter behavior
vim prompts/troubleshooter.txt

Changes take effect immediately (agents reload prompts on each invocation).

Security Considerations

⚠️ Important Security Notes:

  1. .mcp.json executes local scripts - The configuration file runs bash script (scripts/gpt5-agent.sh) when MCP agents are invoked. Always review this script before running.

  2. Supply chain risks - If using npx -y for Context7 or other MCP agents, packages are downloaded and executed automatically. Consider:

    • Pinning specific versions in .mcp.json
    • Reviewing package contents before first use
    • Using local installations instead of npx -y
  3. API keys - Store OpenAI and Context7 API keys in environment variables, never commit them to git.

  4. First-time approval - Claude Code may prompt for approval when first accessing MCP servers. This is expected behavior - review what's being executed before approving.

Recommendations:

  • Review all scripts in scripts/ directory
  • Inspect .mcp.json configuration before use
  • Keep Codex and Claude Code updated
  • Use git to track changes to agent configurations

Key Features

✅ Steering Above All

  • Never repeat guidance - Document once, enforced forever
  • Version-controlled standards - Your guidance evolves with git
  • Automatic enforcement - GPT-5.2 agents apply your rules

✅ Type Safety First

  • NewType for all domain IDs
  • Pydantic schemas for all data
  • Zero Any types in business logic
  • Modern Python 3.10+ syntax

✅ Testability by Design

  • Orchestrator pattern (testable without dependencies)
  • Inheritance-based fakes (no mocks)
  • 90% coverage target
  • Bug fixes require tests

✅ Systematic Debugging

  • 8 troubleshooting principles
  • Search for similar patterns
  • Minimize iteration time
  • GPT-5.2 debugging assistance

✅ Git Worktree Workflow

  • Feature isolation
  • Easy context switching
  • No stashing required

MCP Agent Details

Architecture

Claude Code (You)
    ↓
CLAUDE.md (Workflow guidance)
    ↓
┌──────────────┬──────────────┬──────────────────┬─────────────┐
│              │              │                  │             │
│GPT-Architect │ GPT-Reviewer │GPT-Troubleshooter│  Context7   │
│(Step 0, T-1) │(T, T+1, L-1) │  (When debugging)│(Doc lookup) │
│              │              │                  │             │
└──────┬───────┴──────┬───────┴────────┬─────────┴──────┬──────┘
       │              │                │                │
       ↓              ↓                ↓                ↓
  code-style.md  test-style.md  troubleshooting.md  Library Docs

Agent Capabilities

CapabilityGPT-ArchitectGPT-ReviewerGPT-TroubleshooterContext7
MCP Server Namegpt5-architectgpt5-reviewergpt5-troubleshootercontext7
Read docs-
Read code-
Search codebase-
Library documentation---
API reference lookup---
Usage examples---
High reasoning budget-
Response time5-10 min5-10 min5-10 min<1 sec
ModelGPT-5.2GPT-5.2GPT-5.2Context7 API
Reasoning effortHighHighHighN/A

Contributing

This is a personal steering framework. Fork it and customize for your needs!

To share improvements back:

  1. Fork the repository
  2. Create a feature branch
  3. Update documentation with your pattern
  4. Submit a pull request

License

Apache 2.0 - See LICENSE file

Acknowledgments

  • Claude Code by Anthropic - AI-powered development CLI
  • Codex by OpenAI - MCP server runtime powering the GPT-5.2 agents
  • GPT-5.2 by OpenAI - Powers the three specialist MCP agents (Architect, Reviewer, Troubleshooter)
  • Context7 - Library documentation lookup and API reference

Known Issues

  • Context Rot. As claude does a compaction it forgets most of the instructions in claude.md. Which leads to forgetfulness to call the mcp servers, breaking of style guide, etc. Workaround to the /clear as often as possible, don't schedule two different task into a same context. Split the tasks well.

Collected info

  • 8 stars
  • 1 forks
  • Language: Shell
  • Source updated: 6/8/2026