Section 01
Study Guide
Long-form notes and references.
This study guide provides a detailed review of Claude Code's configuration, mental models, and workflow disciplines. It is designed to help architects and developers understand how to treat Claude Code as a production-grade agent using “Infrastructure as Code” principles.
North Star
Externalize behavior into versioned configuration, separate cognitive capability from execution authority, and keep humans as the final safety controllers for irreversible operations.
Glossary of Key Terms
- CLAUDE.md
- A repository-level instruction file that Claude Code automatically reads each session to understand project conventions and rules.
- Hierarchy (Config)
- The mechanism where Claude loads configurations from the working directory upward to the root, where specific local files override general global ones.
- Plan Mode (--plan)
- An execution mode where the agent proposes a series of steps and tool calls for human approval before performing any actions.
- --print (-p)
- A non-interactive flag used to generate a single-pass output to stdout, ideal for CI/CD environments.
- --allowedTools
- A security flag used to restrict the agent's capabilities to a specific list of tools (e.g., Read only, or specific Bash commands).
- --output-format json
- A flag that produces structured JSON data, including token usage counts, suitable for automated pipeline monitoring.
- --dangerously-skip-permissions
- A flag that bypasses all interactive permission prompts; highly dangerous and forbidden in production/CI/CD pipelines.
- .claudesignore
- A deny-list file that prevents Claude Code from accessing specific files or directories, such as secrets or environment variables.
- Custom Slash Commands
- Reusable, versioned workflows stored in .claude/commands/ that encapsulate specific multi-step tasks.
- Deviation Checkpoint
- A safety gate triggered when the agent identifies a need to change a security-critical decision that was previously approved in a plan.
- MCP Filesystem Provider
- A tool that grants Claude Code read-only access to neighboring repositories, enabling cross-repo awareness without context overcrowding.
- Post-edit Hook
- A configuration that requires the agent to ask for user approval after modifying a file but before running tests or committing.
- HITL (Human-in-the-Loop)
- A recovery pattern where a human interrupts, inspects, rolls back, and re-guides an agent that has drifted mid-task — using Git as the deterministic rollback boundary.
- Source of Truth
- The definitive reference for project knowledge; in this context, CLAUDE.md as the primary guide for agent behavior.
- Context Fork
- A mechanism used in skills to isolate the agent's context, ensuring a clean environment for specific workflows.