Explore, plan, code, commit on existing code

Lesson objective

By the end of this lesson, you will be able to run a non-trivial task on existing code in four steps (explore, plan, code, commit), with plan mode, verification criteria the agent runs and Git commits, and to know when checkpoints are enough and when they are not.

Where it fits

Plan, then delegate with method

How do I run a real task without the agent heading the wrong way, and which tool do I give it for what?

Lessons in this module

1. Explore, plan, code, commit on existing code (this lesson)
2. Subagents, skills, commands and MCP servers: the right tool for each need

What you will learn in the course

This lesson is part of the course Work with Claude Code on a real project: context, planning, review

• Explain what fills Claude Code's context window and manage it (/context, /clear, /compact, resuming and naming sessions), taking cost into account.
• Write a short, useful CLAUDE.md placed at the right level of the hierarchy (user, project, local, subdirectory, path-scoped rules), with imports if needed.
• Run a task on existing code as explore, plan, code, commit, with plan mode, verification criteria and Git rather than checkpoints alone.
• Choose and justify the right mechanism for a need (CLAUDE.md, skill or command, subagent, MCP server, hook) and set it up in the repository.
• Choose a permission mode and write allow, ask and deny rules suited to the project's risks, in the right settings file.
• Write simple hooks as deterministic guardrails (block, format, verify) and know their risks and limits.
• Review the agent's work with a checklist (scope, tests, secrets, migrations, dependencies) and an independent review, and know when to automate it in non-interactive mode or in GitHub Actions.

Related courses

• Build and ship a web app with an AI coding agentJunior · ~2 hr 45 min
• Add a database, authentication and payments to your appAdvanced · ~4 hr
• Git for PMs: ship as a team without putting production at riskAdvanced · ~3 hr 30 min