CLAUDE.md - Specialist Agent

About

Your AI development team. 27+ specialized agents that build, review, debug, and ship production code.

Available packs: Vue 3, React, Next.js, SvelteKit, Angular, Astro, Nuxt

Auto-Dispatch Rules

Automatically delegate based on user intent:

Available Agents

Core Agents

Workflow Agents

Specialist Agents

Support Agents

Available Skills

Security Rules

• NEVER include secrets in shell commands
• Use environment variables for auth
• Reference CI secrets for npm publish

Execution Summary

After each task:

──── Specialist Agent: 2 agents (@builder, @reviewer) · 1 skill (/plan)

Pack Structure

packs/[framework]/
├── agents/           # Full agents
├── agents-lite/      # Lite agents
├── skills/           # Skills
├── ARCHITECTURE.md   # Patterns
└── CLAUDE.md         # Config

Platforms

Hooks

Native Hooks (Claude Code)

Configure in hooks/hooks.json.

Cross-Cutting Concerns

SOLID & Clean Code (Mandatory)

All generated code MUST follow SOLID principles and Clean Code standards:

Clean Code rules: Descriptive naming, small functions (<20 lines), no magic numbers, no deep nesting (max 3 levels), no commented-out code, DRY without premature abstraction.

Observability & Monitoring (Mandatory)

All production code MUST include observability from day 1 - not as an afterthought:

• Structured logging: JSON format, consistent fields (timestamp, level, service, traceId, message)
• Error tracking: Capture errors with context (user, request, stack trace). Never swallow exceptions silently.
• Health checks: Every service exposes /health or equivalent endpoint
• Metrics: Track RED metrics (Rate, Errors, Duration) for critical operations
• Correlation IDs: Propagate trace/request IDs across service boundaries
• Never log: passwords, tokens, PII, credit card numbers, API keys

Agents that MUST enforce observability: @builder, @api, @data, @cloud, @devops, @finance.

Verification Protocol

All agents MUST verify claims with evidence before marking work as complete. No "should work" - run the command, show the output, then claim success. See /verify skill.

Anti-Rationalization

Key agents include rationalization tables that prevent shortcuts. If you catch yourself thinking "just this once" or "it's obvious" - stop and follow the process.

Context Isolation

@orchestrator and @executor use fresh context per subagent. Each task gets a self-contained prompt - no accumulated context pollution.

Anti-Sycophancy

@reviewer agents push back on bad code. No "LGTM" or "Looks good!". Technical evaluation with evidence, YAGNI checks on suggestions.

Persuasion-Backed Enforcement

Discipline agents (@tdd, @debugger, @planner, @executor, @pair) use science-backed enforcement: Authority (IEEE, NASA, Kent Beck), Commitment (contract-based), Social Proof (industry practice).

Governance

• Anti-Rationalization tables in all discipline agents prevent shortcuts
• Verification Protocol ensures claims are backed by evidence
• Execution Summary at end of every task tracks agent/skill usage
• Cost estimation via /estimate before expensive operations
• Health scoring via /health measures project quality over time

Quality Validation

node tests/validate-agents.mjs           # Validate all agents and skills
node tests/validate-agents.mjs --strict  # Treat warnings as errors
node tests/validate-skills-core.mjs      # Test runtime library
node tests/test-skills.mjs               # Behavioral skill tests
node tests/test-skills.mjs --verbose     # With details