HCI OBSERVATORY / INDEPENDENT STUDY

How others study AI interaction: five illustrated readings

Literature and experience review · revision 1 · draft

On this page

An observation series can begin by reading other people’s work. These five primary sources show what was studied, how the argument is written, what the original illustration shows, and what we can learn from it.

Source types: 2 papers, 1 vendor report, 2 firsthand accounts
Scope: Codex desktop, Copilot, a research replay tool, Claude Code and Claude Design
Original illustrations: 5, each credited and dated
Checked: 3 October 2026 · selected readings, not a systematic review

This is a short source review with editorial interpretation. Participants, logs and firsthand experiences belong to the original authors. Different products and dates cannot establish a ranking of today’s Codex and Cowork.

01 · A Codex walkthrough: follow one task#

Author and source: Janine Heinrichs, Unite.AI, 28 May 2026. OpenAI Codex Review.

The author documents a landing-page task through project setup, planning, execution and inspection, with screenshots along the way. This is a personal account, and the publisher discloses affiliate links. It is not a controlled user study.

Janine Heinrichs' Codex screenshot: a Landing Page Test project, an expanded menu with Plan mode enabled and Auto-review displayed below. The author's red rectangle highlights these choices. — Example 1: the author's step-5 screenshot, including the original red annotation. Published 28 May 2026; app build and capture date unverified. © Unite.AI and depicted product owners. Source and context · Full image.

Writing pattern: task → steps → state screenshots → experience. We could add checks for finding the right project and result. A menu image does not verify permission behaviour; the reported speed cannot establish general performance.

02 · Grounded Copilot: build an explanation from behaviour#

Authors and source: Shraddha Barke, Michael B. James and Nadia Polikarpova, OOPSLA 2023. Author-hosted paper · DOI.

The study observed 20 participants and identified acceleration and exploration: using AI to complete a known next step or to discover a direction. Its subject was GitHub Copilot of that period.

Grounded Copilot Figure 1: an input-parsing function with a grey rule.split completion at the cursor, surrounded by the existing code and comments. — Example 2: Figure 1, PDF page 3. The figure region was rendered and cropped, preserving its content; this caption is ours. © Barke, James and Polikarpova, CC BY 4.0. Published April 2023. Original paper.

Writing pattern: observation method → recurring behaviours → explanation → design implications. A question for Codex is whether known and exploratory tasks need different planning support. That transfer requires testing; the paper did not evaluate today’s desktop app.

03 · RECAP: reconnect conversation and code changes#

Authors and source: Keyu He, Qianou Ma, Valerie Chen, Wayne Chi and Tongshuang Wu, arXiv v1, 1 May 2026. Paper and figures.

RECAP links chat and code records. Its university deployment collected records from 41 students, including chat data from 29. The system targets VS Code and Copilot. This preprint’s system demonstration is not a Codex-versus-Claude experiment.

RECAP Figure 2: changed files on the left, code diff in the centre, related chat on the right and a colour-coded event timeline below. — Example 3: RECAP Figure 2, the replay viewer. © He, Ma, Chen, Chi and Wu, CC BY 4.0. Original figure unchanged; source version dated 1 May 2026. Paper and limitations · High-resolution image.

Writing pattern: research question → recording system → linked evidence. Our own task records could connect requests, edits and checks, rather than keeping only the final successful page.

04 · Claude Code logs: define the measures before reading the chart#

Source: Anthropic research team, Agentic coding and persistent returns to expertise, 16 June 2026. Original report.

The report analyses approximately 400,000 Claude Code sessions across CLI, web and desktop channels. People typically determine goals while Claude determines execution. These are classifier-derived transcript findings, not a Cowork button-usability experiment.

Anthropic Figure 2: the x-axis shows Claude's share of decisions per session, and the y-axis shows session frequency. Planning is concentrated at lower shares; execution at higher shares. — Example 4: report Figure 2, unchanged. Blue denotes planning, orange execution. This is a log distribution, not a satisfaction score. Published 16 June 2026; data window October 2025–April 2026. © Anthropic. Data, methods and limits.

Writing pattern: sample → measure definitions → chart → limitations. For Cowork, we can ask whether delegation preserves the person’s goal and review points. Vendor data can inform that question, but cannot establish that a particular layout is better.

05 · A design team’s account: include the process and difficulties#

Authors and source: Tom Moran and Dan Bradshaw, Valliance, 21 April 2026. Claude Design collaborative test drive.

Two designers explored a resource-library interface for about 90 minutes each. They describe different starting approaches, iterations and friction, praising direct visual feedback while questioning mixed chat information. This is a team case involving browser-based Claude Design.

Example 5: the authors' clarification screen in Claude Design, not Cowork. Published 21 April 2026; capture date and build unknown. © Valliance and depicted product owners. Original image unchanged. Full account.

Writing pattern: shared task → different approaches → concrete friction → improvement. We can adopt that structure for feedback and recovery notes. Two authors cannot represent every user, and an April interface cannot establish today’s latest design.

A writing template for this series#

A source review can follow question → author and method → illustration → brief finding → scope → our follow-up question. Add our own participants, task, platform and results only when an actual study happens.

Source review: read and explain existing work; a new experiment is not a prerequisite.
Interface reading: explain a sourced screen, separating facts from judgement.
Firsthand observation: document our actual task and screenshots.

All three belong in the Observatory. Papers, vendor reports and personal accounts have different uses; inspect how evidence was obtained before extending a conclusion.

Sources and revision#

This page includes limited illustrated excerpts and short summaries; follow each author link for the full text. The research image register links every image to its origin. Both papers carry CC BY 4.0; other images retain their owners’ credits. Retrieved 3 October 2026. The theme licence does not cover these third-party materials.

Revision 1 · 3 October 2026: five primary sources checked, five examples included, with writing patterns and scope. AI-assisted draft awaiting personal review.

Codex source review and interface reading → · Claude source review and interface reading → · Method →

← All observationsBack to top ↑