Covers install — one command puts the claude binary on PATH, no model weights download locally. Authentication trades an API key or console login for a persistent session on first run. Permission modes (default, acceptEdits, bypassPermissions, plan) trade autonomy for safety. claude.ai/code runs the same agent in a browser tab with no local install required, and shared sessions let work started in web continue in terminal and vice versa.

Skills are versioned slash-command files Claude Code recognizes from .claude/commands. Each Markdown file becomes a command — type /name and the agent loads the prompt, then runs it. Frontmatter declares an allowed-tools list and a description; the body uses Bash blocks for read-only inspection and an Instructions section for what to change. Skills make recurring workflows discoverable and reproducible across the team.

Claude Code's Agent tool spawns a fresh sub-agent in an isolated context to run a focused job, then returns a structured report. Use it to dispatch read-only reviews in parallel against a contract, keeping the main session lean. Sub-agents make complex tasks reliable: divide work, scope tools, then merge findings. This module covers when to delegate, how to write contracts the sub-agents must satisfy, and how parallel review increases coverage without explosion of complexity.

Production code needs more than agents. This module locks the loop with pmat — a Rust CLI that replaces grep with ranked function queries by Technical Debt Grade, complexity, and coverage — plus contract gates that validate every change and sub-agent output. CLAUDE.md routes the agent to pmat for search, pmat hooks install enforces pre-commit gates, and YAML invariants pin sub-agent shapes. The full claude-from-zero repo demonstrates all four surfaces working as one cycle.

Synthesize the full Claude stack — surfaces, tool surface, skills, sub-agents, and pmat contracts — into one repeatable, gated workflow you can take into your own repo.