Project: review Harbor booking access across states

Last updated: 7 Oct 202618 min read

project

A prompt can organize a review of observable interface states. It cannot infer keyboard behavior, spoken output, or conformance from a still image. Record the tested build and state, inspect native semantics and actual interactions, then send reproducible findings to an owner. Harbor Dispatch is a fictional booking interface used throughout these lessons.

Review the packet

Harbor Dispatch has a booking form, a confirmation dialog, and 47 queued jobs in its test dataset. The reviewer records build B-47, uses a keyboard to open and close the dialog, tests a failed form submission, and checks a narrow viewport. The initial finding is specific: after closing the dialog with Escape, focus moves to the page top rather than returning to the button that opened it. A screenshot alone could not establish that failure. The release decision waits for a repair and a repeat of the same steps.

Build B-47 | booking form | test account without private data
Open dialog with keyboard; close with Escape
Observed: focus at page top
Expected: focus returns to opening button
Finding: reproducible; owner and retest required

Performance and review cost

Walking N focusable controls is O(N) interaction steps, and recording S meaningful interface states adds O(S) review records. Automated checks can scan markup quickly, but keyboard, zoom, and spoken feedback still require human verification. Save only the evidence needed to reproduce a finding; an unbounded transcript of every control makes the handoff slower to inspect.

Common Mistakes

Do not claim accessibility from a screenshot alone.
Do not treat a model's suggested fix as a verified repair.
Do not release a finding without the tested build and interaction steps.

Related lessons

prompt engineering

accessibility review