Complete Domain Mastery

Domain 3: Orchestrating Claude Code Workflows

Domain 3 defines the operational framework for Claude Code. It moves beyond "chatting" and into professional orchestration: where context is managed spatially, capabilities are extended programmatically, and quality is enforced through iterative cycles and automated pipelines.

I. Spatial Context & Configuration Hierarchy

Hierarchy & Scoping

The exam focuses heavily on how Claude resolves conflicting instructions across different file system levels. Understanding the Distance-Based Priority is non-negotiable.

Tasks 3.1 & 3.3 Mastery

📂 Environmental Inheritance

Claude builds its worldview by merging instructions from multiple layers. The "Winner" is always the rule closest to the active file.

1. Global Layer (~/.claude/CLAUDE.md) Personal productivity hacks, preferred editor settings, and cross-project safety constraints. This is the background persona.

2. Project Layer (root/CLAUDE.md) The "Source of Truth" for the repo. Contains core build commands, language versions, and architectural mandates that apply to the whole team.

3. Path Layer (subdir/CLAUDE.md) Context-specific overrides. Essential for monorepos (Frontend vs. Backend) or security-critical folders (Infra/Auth).

Strategic Application

🔍 Scoping & Context Pollution

Architects must minimize "Context Pollution"—situations where Claude uses the rules of Module A to solve a ticket in Module B.

The "Monorepo Leak" Fix If Claude tries to use `pytest` for a React file, you have a scoping problem. The solution is creating a local `/web/CLAUDE.md` to trap the context.

Selective Overriding Subdirectory rules are additive. They don't erase the root; they only replace specific headings or bullets. This preserves general standards while allowing local flexibility.

II. Skill Extensibility & Multi-Tool Orchestration

Extensibility

Task 3.2 Mastery

📜 Transformative Skills

Beyond built-in tools, Domain 3 covers how to extend Claude's utility through custom interfaces. This is where Claude evolves from an Assistant to a specialized Agent.

Slash Commands (/command) Low-complexity, high-repeatability shortcuts. Ideal for common developer tasks like /lint, /test-file, or /gen-docs. These act as macros for long, multi-step prompts.

Agent SDK Skills High-complexity capabilities. Skills allow Claude to perform complex orchestration, handle state across turns, and interact with multiple MCP servers as a single cohesive unit.

Expert Domain Encapsulation By defining skills in `CLAUDE.md`, you are effectively "injecting" senior-level domain expertise into the LLM's working memory for every project contributor.

III. Execution Strategy & Quality Cycles

Strategic Execution

The Architectural Decision Matrix

Task Characteristic	Mode Selection	Refinement Strategy	Exam Success Point
Unfamiliar Legacy Repo	Plan Mode (`-p`)	Discovery-first. Analyze hierarchy before making ANY changes.	Prioritize "Plan Proposal" over immediate "Direct Edits".
Known Single-File Bug	Direct Execution	Direct modification (hammer mode). Quick iterative feedback.	Avoid over-planning simple tasks to save time/tokens.
Core Auth/DB Refactor	Plan → Direct	Layered Refinement. Refine the core logic first, then security, then style.	Identify "One-Shot" attempts as an anti-pattern for high-risk modules.
Multi-Turn Code Loop	Any	TDR (Test-Driven Refinement). Use test failures as the objective signal for turns.	Feed objective machine signals (tests/lints) over vague prose feedback.

IV. Automation & Governance Pipelines

Continuous Improvement

🚀 Non-Interactive Mode

For CI/CD (GitHub Actions, GitLab), the --yes flag is mandatory. It bypasses the "Human-in-the-loop" requirement for cost and safety approvals.

🔒 Credential Safety

Secrets (`ANTHROPIC_API_KEY`) must be masked. Automated Claude reviewers should have scoped tokens and restricted filesystem permissions in the runner.

✅ Automated Synthesis

The ideal pipeline uses Claude to review PR diffs, check them against the local `CLAUDE.md` standards, and fail the build if high-severity violations occur.

🎯 950+ Architect's Subtle Nuance

The "Global Leak" Paradox

A common 950+ score trap: putting project-specific rules (like "use pnpm") in your Global ~/.claude/CLAUDE.md. While it works for you, it won't work for the CI/CD pipeline or your teammates. This causes "Works on my machine" failures in Claude Code. Global is for personal style (e.g., "be concise"); Project Root is for shared technical truth.

The "Non-Interactive" CI Hang

In automated environments, Claude Code will hang indefinitely if it hits a confirmation prompt (like "May I run this command?"). To score 950+, you must know that only the -p (or --print) flag combined with --yes ensures a zero-hang pipeline. Vague instructions like "don't ask for permission" in the prompt have a non-zero failure rate.