Claude Code Workshop: From First Prompt to Production-Ready Pull Request

Claude Code is a coding agent that works in your terminal and project directory. It reads files, proposes and makes edits, runs commands and iterates on results. That power makes habits matter: the same tool can produce a clean, reviewed pull request or a sprawling change nobody understands. This workshop focuses on the habits.

The path from prompt to pull request

We structure the day around one repeatable loop, practiced several times on increasingly realistic tasks.

1. Frame the task. Write a short brief: goal, constraints, files likely involved, and how you will know it works.
2. Ask for a plan. Review the plan before any edit happens. Correct wrong assumptions early, when they are cheap.
3. Implement in small steps. One step, one verification. Run the tests or the app after each step.
4. Review the diff yourself. Read it as if a colleague wrote it. Question anything unexpected.
5. Prepare the pull request. Clear description, linked ticket, notes on what was tested and what was not.

Topics covered

• Installing and configuring the tool, and understanding what it can access on your machine
• Writing and maintaining a project instruction file such as CLAUDE.md (our guide on AGENTS.md and CLAUDE.md goes deeper)
• Permissions: what to allow, what to confirm each time and what to forbid
• Managing context so the agent stays focused during long sessions
• Using the agent for exploration, refactoring, debugging and test writing
• Connecting external tools carefully, see MCP servers for coding agents

Common mistakes we address

Participants usually recognize at least one of these from their own experience: asking for too much at once, accepting a summary instead of checking the result, letting the agent fix a failing test by weakening it, and skipping version control checkpoints. We create small traps for each of these in the exercises so the lesson is felt, not just heard.

A worked example

To make the loop concrete, imagine a task such as adding input validation to a sign-up form and returning clear error messages. In the workshop, participants first ask the agent to read the relevant files and explain the current flow without changing anything. Next they request a plan listing the files to edit and the tests to add. Only after correcting the plan do they allow edits, and even then in two steps: validation logic first, user-facing messages second.

After each step they run the test suite and open the form in the browser. The instructor then introduces a twist, for example an existing test that fails for an unrelated reason, to observe whether the agent fixes the cause or merely makes the test pass. Discussing that moment is usually the best part of the session, because it shows why the person reviewing the diff needs to understand the intent, not just the green checkmark.

Take-home material

Participants receive a starter instruction file, a permission checklist and a pull request template that asks the author to state what the agent did, what the human verified and what remains untested.

Where human responsibility fits

At apprat.io the agent does much of the implementation, while a senior AI Code Engineer plans the work, reviews every change, tests on real devices and takes accountability for the outcome. That is the model instructors teach, because it is the one we use for client projects. Features that touch authentication, payments or personal data get extra scrutiny regardless of how confident the agent sounds.

Who should attend

The workshop suits developers comfortable with Git and a terminal, and tech leads who want to set team standards. Pairing it with a team-wide day, as described in our guide to agentic coding workshops for teams, works well. For rollout after the workshop, see the 30-day Claude Code onboarding plan.

Check Anthropic's current documentation for installation, pricing and supported platforms, since these details change. If you would like a session tailored to your stack, you can book a call for free.