Claude Code Guide: Agentic Engineering Workflow

Learn to understand an unknown repo by cloning a GitHub project, asking clear questions, and using Cloud Code to explain the code base and provide a concise overview.

Explore an optional exercise that prompts you to ask three questions about a repo, clarifying the language, a skill, and a hook, recognizing markdown as the content format.

Explore how models, tokens, and the context window drive Cloud Code usage, with practical guidance on resetting context, choosing Opus, Sonnet, or Haiku, and managing plan limits.

contrast vibe coding, where code is treated as ai's problem and lacks documentation, with agentic engineering that requires planning, documentation, tests, and code reviews to establish conscious decisions.

Learn how to use CloudMD to guide agent behavior with per-project and local overrides, manage multiple CloudMD files, and apply concise conventions for the generated code.

Learn to craft a project-specific CloudMD file by merging Karpathy’s guidance with practical CloudCode insights, using init to analyze code, emphasize TypeScript, and test with unit tests rather than mocks.

Explore how Claude integrates Git to track changes, stage, commit, push, craft commit messages, and manage rebases and stashing, then leverage Git worktrees to work on multiple branches simultaneously.

Initialize the repository, stage files, and commit the first changes. Create a remote on GitHub, set up SSH keys, and push, while using Cloud to run Git commands.

Configure tool permission rules in cloud code to whitelist safe commands, deny sensitive access, and require prompts; manage per-project settings via the UI or .cloud/settings.json.

Learn how to set up a remote repository on GitHub, choose between SSH or HTTPS, and push your code while managing git ignore and co-author settings.

Explore plan mode in cloud code to convert coding tasks into a validated plan for pdf pipeline, parse invoices into sqlite, and produce top five vendors by total spend.

Build an automatic code-review sub-agent that focuses on security, best practices, and readable, reusable code; triggers after a chunk of work and only reviews uncommitted code, without running tests.

Learn to create and reuse agent skills with markdown files and terminal commands for on-demand review. Compare skills with sub-agents, where skills describe tasks and sub-agents execute them.

Create a custom code-review skill for Cloud Code that runs on demand with the CodeReviewer sub-agent, reviewing uncommitted changes or specified files, stored under .cloud/skills, and renamed to CodeReview.

Explore the skills.sh marketplace by Vercel to find and install skills via npx, ensure Node.js installed, and review security audits to avoid prompt injection and unmaintained tools.

Learn how hooks enforce non negotiable rules in Cloud Code, reacting to events before or after tools, protecting .env secrets, tests, and code formatting by managing context and prompts.

Build hooks for cloud code: run tests after stop and block edits to invoices CSV, then enforce dead code removal and formatting with hard stop protection for the secrets file.

Discover MCP, the model-context-protocol that lets CLOD code call typed tools instead of shell commands. See how MCP servers streamline integrations with services like Postgres and GitHub.

Learn practical MCP with Playwright by installing a Playwright MCP server to control a real Chrome session against the MoneyPools app. Validate front-end behavior with end-to-end tests and page descriptions.

Install the superpowers plugin for the Money Pools project, explore plugin scopes, run a hands-on design and verification flow with cloud code, Playwright MCP, and auto-generated design docs.

Explore cloud code permission modes, from default least permissive to accept edits, plan mode, auto mode, classifier, and bypass. Start with least permissive settings and advance only after pre-approval.

Understand bypass permission mode in Cloud Code, its risks, prompts, and how it bypasses project permissions. Use container isolation or virtual machines to contain it and guard against prompt injection.

Learn to use cloud code git worktrees to run two Claude agents on parallel worktrees, develop separate features concurrently, test independently with SQLite, and merge changes into main.