AI Coding Mastery

Install Claude Code (or Cursor) and launch it inside a fresh, empty project folder. Ask it to create a file that prints your name. You are done when the agent created the file and you ran it successfully.

Write one clear prompt with a goal, constraints, and a check, and have the agent build a working single-page to-do list. Done when you can add, complete, and delete tasks in your browser.

Put your project under Git and commit. Then deliberately ask the agent to make a change you do not want, review the diff, and use git restore to roll it back. Done when your project returns to the committed state.

Have the agent build a small one-page site about you or a hobby, then deploy it to Vercel, Netlify, or GitHub Pages. Done when you can open the live URL on your phone.

Find a friend or family member who does not code and explain, in under two minutes and no jargon, what a coding agent does (read, act, check, repeat). Done when they can repeat the four steps back to you.

On a small real repo, fix or refactor something by pointing the agent at only the two or three files that matter (using @ or paths), then /clear and start a second unrelated task. Done when both tasks succeed and you never pasted a full file.

Run /init or write a CLAUDE.md (or .cursor/rules) for a real project with your commands, conventions, and a 'do not' list. Done when a fresh agent session follows at least one rule you set without being reminded.

Pick a real feature, have the agent produce a plan first, push back on at least one point, then let it implement after approval. Done when the merged change matches the approved plan and tests pass.

Add one MCP server (GitHub, Playwright, or Postgres) to your agent and complete a task that requires it, such as fixing a logged issue or querying the database. Done when the agent successfully calls an MCP tool to finish the work.

Without a screen, sketch your real workflow on paper: where you point context, where you plan, where tests gate completion, and where MCP tools plug in. Done when you can show someone the diagram and explain each guardrail.

Create a reviewer subagent (a Markdown agent file with scoped tools and a cheaper model) and run it against a recent diff. Done when it returns at least one specific, file-and-line finding on real code.

Add a Stop hook that runs your test and type-check commands and blocks completion on failure. Done when you watch the agent reopen a task because a failing hook refused to let it finish.

Build a small MCP server exposing one tool that wraps a real API or script, connect it to your agent, and complete a task that requires it. Done when the agent calls your tool correctly and uses the result.

Wire a headless agent into a pipeline or cron job (for example, triage a labeled issue into a draft PR) with pinned model and allowed tools. Done when a trigger produces an artifact (PR or report) with no human in the loop until merge.

Take a real agent-written diff to a colleague (or a developer friend) and review it together away from your usual setup. Done when you have agreed on at least two findings and a rule to add to your memory file or hooks.