Convenient for whom?

What counts as an input device?#

Vision Pro includes sensing hardware as well as displays; gaze-and-pinch is the input technique analysed in HW2. ChatGPT and Claude are software interfaces; conversational instructions are an interaction technique. Codex additionally supports delegating work on files. Douyin and Xiaohongshu are applications combining graphical controls, content discovery and recommendation. Calling every whole product an “input device” would blur useful distinctions.

For HW2’s new-or-failed-technique requirement, a specific contemporary technique is easier to defend than a review of a long-established content app. The assignment permits a technique as well as a physical device.

Codex · Delegating work and reviewing the result#

Goal. Change a website and understand what changed before publishing it.

Analysis. Official documentation describes reviewing code changes and pointing to elements in a browser to request edits. These offer complementary ways to connect an instruction to its result. Code review · Browser annotations.

Natural language reduces the need to remember command syntax. Pointing at the actual paragraph can also be clearer than describing its position. However, a novice may be able to request a change without being able to judge a code diff. A polished preview confirms appearance, not correctness of every behaviour. Delegation shifts some effort from doing the work to checking the work.

Redesign. Pair technical changes with a short explanation of the visible behaviour, a preview, the checks actually performed, and a clear statement of whether the result is local or published. This website workflow is a good context in which to evaluate that distinction. These are proposed design priorities, not a claim that the product lacks all of these facilities.

ChatGPT · Easy questions, harder verification#

Goal. Understand an unfamiliar assignment and trace factual claims to useful evidence.

Analysis. Conversation allows a learner to ask incomplete questions and refine them. Official documentation says web search can surface sources and citations in the conversation; availability can depend on workspace settings. Web search.

The hypothesis is that follow-up dialogue lowers the effort of getting an initial explanation. The corresponding risk is that fluent prose can feel more certain than its evidence warrants. A link is useful only if it supports the associated claim. Asking for a shorter answer may improve readability while removing qualifications the learner needs.

Redesign. Put “what the assignment explicitly requires” next to “suggested writing structure.” Distinguish source facts from interpretations, and offer a short explanation with expandable detail. To evaluate this, test whether learners can identify both the required deliverable and a claim that still needs checking; satisfaction alone is insufficient.

Claude · An editable artifact beside the conversation#

Goal. Create an output that can be inspected and revised without searching a long chat transcript.

Analysis. Claude’s documentation describes artifacts such as documents, designs and interactive tools opening beside a conversation, where they can be revisited and edited. This analysis concerns that artifact workflow, not every Claude product or a model-quality benchmark. Artifacts documentation.

A persistent visible object can make it easier to refer to “this paragraph” or “that chart.” The design risk is version ambiguity: the user needs to know whether a message proposed a revision or actually changed the artifact, and which copy will be shared. This is a general question to test in the workflow, not an observed failure in a specified Claude release.

Redesign. Emphasise the current version, the exact scope of a revision and the distinction between saving, exporting and sharing. Compare how reliably users recover an earlier wording after several revisions. The lesson also applies to other chat-plus-document systems, including this website’s editing workflow.

Douyin · Recommendations and intentional control#

Goal. Find relevant entertainment while retaining control over the session.

Analysis. Douyin’s privacy policy describes personalised recommendations informed by activity such as browsing, searches and interactions, and documents controls for managing personalisation. The existence of these controls should be acknowledged before proposing improvements. Official policy, sections 1.5.1 and 3.1.5.

Recommendations can reduce repeated searching. However, a behavioural signal is not a perfect statement of intent: spending time on a clip could indicate curiosity, confusion or disagreement. A useful HCI question is whether users can readily correct an unwanted inference. Another is whether finding appealing content helps or conflicts with the user’s intended session length. Neither concern establishes that every recommendation is deceptive or that every user loses control.

Redesign. Test a nearby explanation of “why this appeared,” an understandable correction control and optional stopping cues. This is a proposal to evaluate discoverability and usefulness, not a claim that no related controls exist. Measure success against the user’s stated goal, including the ability to leave when intended, rather than watch time alone.

Xiaohongshu · Discovery is not the same as decision quality#

Goal. Find useful experience-based information without mistaking popularity for suitability.

Analysis. The platform’s published “hot topics” rules describe personalised topic suggestions and an option to disable personalisation. This is evidence about that feature; it does not verify the layout of every current search or recommendation screen. Official feature rules.

For a task such as choosing a travel destination, the evaluation should ask whether the user can find relevant constraints: location, date, budget and whose experience is described. Attractive presentation can help someone notice a post, but it does not establish that its advice fits their circumstances. Popularity and recommendation are different signals from verified relevance.

Redesign. Make decision-relevant details and commercial relationships easy to inspect, and let readers compare posts against their own criteria. This is a proposed task-oriented design, not evidence that particular posts are undisclosed advertisements. A later walkthrough should use a specified app version and dated screenshots before making claims about button placement or missing labels.

Vision Pro · A separate, focused HW2 case#

Read the gaze-and-pinch analysis. Its key question is how the system distinguishes looking at something from deciding to activate it. A small gesture may work well for a discrete selection while a keyboard remains a more suitable comparison for sustained text entry. “Newer” is not itself evidence of better task performance.

How to make this critique stronger#

Choose one task for each comparison, specify the version, and record the interaction with a screenshot or short video. Compare completion, errors, recovery and user control. State who performed the test and do not invent personal experience. The goal is an explainable judgement supported by observations, not a numerical league table without data.