AI Linkbase
Operator pathLesson 03

Use a Coding Agent on a Real Project

Use a coding agent to inspect, change, and verify a small project while keeping every modification reviewable and reversible.

Time
45 minutes
Level
Beginner to intermediate
Reviewed
August 1, 2026

Lesson outcome

You will complete one scoped code change with a baseline, an explicit plan, a readable diff, and verification evidence.

Before you start

  • A Git repository or disposable practice project
  • A coding agent such as Codex, Claude Code, Cursor, or Copilot

Mental model

Three ideas to keep

Baseline before edits

Record the branch, existing uncommitted changes, and current test result so new failures can be separated from old ones.

Scope is a safety tool

Name the files and behavior in scope. Tell the agent to preserve unrelated changes and ask before expanding the task.

Evidence closes the loop

A code change is not complete when the files look plausible. It is complete when the relevant build, tests, and user flow are verified.

Follow along

Complete the workflow

  1. 01

    Capture the baseline

    Run the repository's status command and its smallest relevant test. Save the output before asking for edits.

    Expected result: You know which changes and failures already existed.
  2. 02

    Write a bounded request

    Describe the intended behavior, files or area in scope, acceptance criteria, and actions that require approval.

    Expected result: The request can be checked without guessing what 'better' means.
  3. 03

    Ask for inspection and a plan

    Let the agent read the relevant code and explain the change path before editing. Correct the plan if it touches unrelated systems.

    Expected result: The plan names the affected route, component, data source, and verification step.
  4. 04

    Review the diff

    Inspect the changed files. Look for deleted behavior, broad refactors, new dependencies, secrets, and unrelated formatting churn.

    Expected result: Every changed line can be connected to the requested behavior.
  5. 05

    Verify the story

    Run the focused test, production build, and the actual browser flow when UI behavior changed.

    Expected result: You have test output and a visible result, not only the agent's summary.

Practice task

Make one reversible UI improvement

Ask the coding agent to add a small empty-state message, improve one label, or fix one responsive spacing issue in a practice project.

You are done when:

  • The initial Git state was recorded.
  • Only expected files changed.
  • The build passes and the page was checked at desktop and mobile widths.

Guardrails

Keep the exercise safe and reviewable

  • Do not let an agent discard pre-existing uncommitted work.
  • Do not paste production secrets into prompts or test fixtures.
  • Require explicit approval for deployments, database writes, and destructive Git operations.