GitHub CLI
The first step is making sure your code agent can interact with GitHub. The
most direct way is the GitHub CLI (gh).
"Help me install the GitHub CLI and authenticate with my GitHub account."
Optionally, install the GitHub MCP in your tool. It gives the agent native access to issues and PRs. But if it doesn't work, move on. The CLI is enough.
Run gh auth status in your terminal. Your account should show as
authenticated.
Create the repository
"Create a repo for me."
"Create a public repository on GitHub called 'deezer-explorer' with the description 'Site for exploring artists and discographies via Deezer API'. Clone it locally and open the folder in my development tool."
Run gh repo view from inside the folder. The repository info should
appear.
Harness engineering
Harness engineering means preparing the environment where the agent works: context, plan, rules, and quality criteria. It is not busywork. It keeps the AI from making things up.
Before asking for code, ask for the files that make the project readable for humans and for AI.
"Create a README.md for this project. Include: site goal, audience, expected features, how to run locally, how to test, how to publish on GitHub Pages, and known limitations. Use simple language, as if someone were joining the project now."
"Create an initial PLAN.md to implement the project in small phases. Each phase needs a goal, testable acceptance criteria, expected manual validation, and open decisions. Include known risks, such as CORS in the Deezer API. Do not implement anything yet."
"Create an AGENTS.md file at the project root with: overview (static site consuming the Deezer API, published on GitHub Pages), sections for project commands, quality standards, how to update PLAN.md when decisions change, and operating rules for the agent. If I'm using Claude Code, create CLAUDE.md for that same role."
Read all three files. If there is a command that does not exist, invented scope, vague promises, or a technical decision you do not understand, ask for an adjustment. A good file helps you hold the agent accountable later.
README.md, initial PLAN.md, and AGENTS.md (or CLAUDE.md in Claude Code) exist, have been read, and make sense to you.
Chapter 1
0 of 3 checkpointsComplete all checkpoints to unlock the next chapter.
Next up: Plan with the agent →