This Claude Code Toolkit Replaces Generic Prompts with 60+ Specialized Agents

Most Claude Code setups are just thin wrappers around "You are an expert at..." prompts. The Claude Code Toolkit takes a fundamentally different approach: instead of one general-purpose model, you get a router that automatically selects specialized agents and structured workflows for every task.

The Architecture: Router → Agent → Skill → Script

The toolkit introduces a /do command that acts as a single entry point for everything. You describe what you want, and the system:

1. Routes your request by domain and action
2. Selects a specialized agent (60+ available)
3. Loads a workflow skill (115+ available) to enforce methodology
4. Uses Python scripts for deterministic validation

Example: /do debug this Go test automatically routes to:

• golang-general-engineer agent (Go-specific context)
• systematic-debugging skill (enforces REPRODUCE, ISOLATE, IDENTIFY, FIX, VERIFY phases)
• go-testing scripts for validation

Each phase has a gate, ensuring structured progress without endless back-and-forth.

Installation in 3 Commands

git clone https://github.com/notque/claude-code-toolkit.git ~/claude-code-toolkit
cd ~/claude-code-toolkit
./install.sh --symlink  # recommended for updates

Then start Claude Code and run:

claude
/install

The installer places everything in ~/.claude/ and configures hooks in settings.json. Use --symlink mode to update with git pull, or --copy for a stable snapshot.

Warning: Symlink mode replaces directories—back up existing customizations first.

Force Routing for Critical Domains

The toolkit includes force-routing overrides for domains where generic advice fails:

• golang-concurrency → golang-concurrency-engineer + concurrency-systematic
• golang-error-handling → golang-error-engineer + error-wrapping-systematic
• python-async → python-async-engineer + async-systematic

These exist because generic advice about Go concurrency or error wrapping often produces broken code. The specialized agents know domain-specific idioms and patterns.

Comprehensive Code Review with 20+ Specialists

The /comprehensive-review command runs a three-wave review process:

Wave 0: Package discovery with per-package language specialists

Wave 1: Parallel foundation review (security, concurrency, silent failures, performance, dead code, API contracts, naming consistency)

Wave 2: Cross-cutting analysis where reviewers read all Wave 1 findings

Example: If a security reviewer flags a swallowed error in Wave 1, the concurrency reviewer in Wave 2 can trace that error to a concurrent code path.

Result: Unified BLOCK/FIX/APPROVE verdict with severity ranking.

Automatic Learning Capture

Hooks automatically capture learnings into a SQLite database (learning.db):

• error-learner: Captures tool errors and their solutions
• review-capture: Extracts findings from reviewer agents
• session-context: Loads high-confidence patterns at session start

On your next session, the retro-knowledge-injector hook queries the database via FTS5 search and injects relevant patterns, creating a feedback loop that improves over time.

When This Toolkit Shines

Use this when:

• You're tired of writing the same prompt patterns repeatedly
• You need consistent methodology for complex tasks (debugging, refactoring)
• You want domain-specific expertise without manually switching contexts
• You work across multiple languages and need specialized tooling for each

Skip this if:

• You prefer manual control over every agent selection
• You have a minimal Claude Code setup you don't want to modify
• You only work in one language with simple tasks

Maintenance and Updates

In symlink mode, update with:

cd ~/claude-code-toolkit && git pull

Run /install again anytime to verify installation or diagnose issues. The health check shows what's available and walks through first-time setup.

The toolkit represents a year of daily use refinement—not just another prompt wrapper, but a complete system for structured, domain-aware development with Claude Code.