Claude Code: Plan, Build, Then Review a Small Change

Anthropic recommends exploring before coding and providing observable verification. The workflow here applies that idea to a small label fix. Approval of a plan is not approval to deploy, and a passing test does not substitute for reading the patch.

Complete a read-only inventory and confirm the project’s instructions. Preserve existing changes. Define an allowed file set and a test that would fail on the current form. Do not assume a particular planning-mode button exists in every client.

Plan: add a visible label linked to the email input, reuse the existing form spacing, and leave submission logic unchanged. Before editing, type an address and observe that the placeholder disappears. Record this as the failing user-visible check, not as a successful label test.

• Patch boundary: contact markup plus existing form CSS only.
• Review question: does the label’s for value match the input id?
• Regression question: is the original submit handler untouched?

1. Ask for a plan listing exact files, intended behavior, checks, and unknowns. Resolve unknown ownership before editing.
2. Approve the bounded change and require the agent to preserve unrelated work.
3. Run the known project checks and repeat the browser observation with text entered. Tab through all controls.
4. Read the diff separately from the completion summary. Ask for a correction if the patch introduces unrelated cleanup or weakens a test.

A good result is a small patch plus evidence of the original failure and the corrected behavior. Report checks with their real outcome and environment. A browser-only check should not be described as end-to-end delivery verification.

Stop if the plan unexpectedly needs a framework migration, if the baseline test cannot run, or if the agent removes an assertion to make tests pass. Separate an environment failure from a code defect before proceeding.

• Did the acceptance check actually fail before the change?
• Did only approved files change?
• Does the final report include limitations and the exact remaining untested behavior?