A polished event page can hide a form that never tells the visitor what happened. The spacing looks right. The cards line up. The blue button sits exactly where the designer wanted it. None of that establishes whether someone can find a workshop, correct a mistake and finish a request for information.
The useful review unit is a journey through states. A screenshot records one of those states, often the one with the least uncertainty. A task walkthrough records the action that produced it, the feedback that followed and the way out when the input was wrong. That changes what you ask an AI assistant to build, and what evidence you accept when it says the work is finished.
Consider a fictional repair workshop page. It lists a lamp repair session on November 14 and a bicycle repair session on November 21. A date filter narrows the list. A request-information form takes an email address and a question. This is a synthetic local demonstration with no live message delivery, registration or customer data. The boundary matters most at the point where the interface looks successful.
The visitor's task is specific: find the November 21 workshop, ask whether a bicycle with a damaged wheel is suitable, and understand whether the request actually went anywhere. Before implementation, the owner needs to approve that journey. An assistant should not quietly turn an information request into a booking, add a payment step or invent a promise that someone will reply within a day.
Give the assistant the two invented events and the expected transitions. The initial state shows both workshops. Selecting November 21 shows the bicycle session alone. Selecting November 28 shows no matching workshop and a way to clear the filter. Clearing it restores both events. Those are separate acceptance conditions, even when all four views share the same attractive layout.
The empty state deserves its own sentence. A blank area can mean there are no matching events, the data failed to load or the interface broke. In this local example there is no remote data source, so the intended message can be precise: no workshops match this date; clear the filter to see all workshops. A production version with remote data would need a separate failure state rather than reusing that explanation for a network error.
The form creates a different set of transitions. Submitting it with an empty email should leave the visitor on the form and identify the missing information. Correcting the email while leaving the question blank should still prevent completion. The question already entered should survive an email correction. A visitor should not have to reconstruct their request because one field failed validation.
W3C's form notification guidance calls for feedback about both successful and unsuccessful submissions, with clear errors and instructions for resolving them. That supports a concrete review question: can the visitor identify what needs fixing and continue? A red outline alone does not answer it. The reviewer needs to trigger the error, read the explanation and follow the correction path.
Discussion Topic:
Which state does your website review miss most often: empty results, invalid input, or recovery?
A valid local submission should end with an honest result: the demonstration accepted the input, and nothing was sent. Calling that state “message delivered” would claim an effect the system cannot produce. The distinction belongs on the page itself. A reviewer who knows the implementation is a mock is not a substitute for a visitor receiving accurate feedback.
There is a reasonable argument for starting with pictures. Screenshots make visual regressions visible and let a team compare spacing, contrast and hierarchy without replaying every interaction. Keep them. The mistake is letting evidence for appearance answer a question about behavior. A screenshot of a success message cannot establish that validation ran, that the correct input reached it or that an external service received anything.
Use a compact walkthrough record beside the implementation. Each entry needs a starting state, an action, an expected result and an observed result. Add the build identifier and review conditions once at the top. Leave the observed result blank until someone performs the action. An assistant can draft the expected sequence, but an expectation copied into the observation column is still an expectation.
The worked sequence for this workshop is short enough to inspect without a test framework. Start with two events, choose November 28 and expect an explicit empty result. Clear the filter and expect both events again. Choose November 21 and expect only the bicycle workshop. Submit an empty form, expect a correction message, enter a fictional address and question, then expect a local confirmation that says nothing was sent. Each transition supplies a distinct place where the journey can fail.
Repeat the journey with the keyboard. WCAG's keyboard criterion requires functionality to be operable through a keyboard interface, with an exception for functions that depend on a movement path. Filtering events and entering a question are ordinary controls, not freehand drawing. The review should establish that focus reaches them, their actions work and the visitor can continue after an error. A screenshot cannot record that sequence.
Small-screen review also needs the error and empty states. A narrow initial view may look tidy while a longer validation message pushes the submit button outside its container. W3C's reflow guidance describes presenting vertically scrolling content at a width equivalent to 320 CSS pixels without losing information or functionality or requiring two-dimensional scrolling, subject to stated exceptions. Use that as a concrete check, not a claim that one viewport establishes accessibility compliance.
The assistant's completion report should distinguish implementation from observation. “Added a date filter” describes a change. “Selected November 28 and observed the empty message on this build” describes evidence. If nobody has exercised keyboard operation or inspected the narrow layout, those checks remain pending. A generated walkthrough helps organize the work; it does not execute itself.
Keep the acceptance decision bounded to the demonstration. Local validation cannot establish email delivery, server-side rejection, spam handling or behavior after deployment. Those require a separately scoped implementation and new evidence. The owner can approve the local journey while withholding permission to release it publicly. That is a useful intermediate result, provided the report names its limits.
When the next website preview arrives, ask for the path from an ordinary mistake back to a completed task. The most revealing frame may be the one nobody chose for the screenshot.
If this was useful, forward it to one engineer who needs less noise in their feed.


