Research Context
AI capabilities have moved fast. Interface design has not. Every major platform still defaults to a scrolling text thread — a format designed for conversation, now being used for planning, execution, delegation, and supervision. This archive documents what a better alternative looks like, pattern by pattern.
Most AI products are using the same interface pattern — a scrolling text thread — regardless of whether the task is a question, a workflow, a delegation, or a collaboration. This archive is built on the premise that this is a design mistake, not an inevitable consequence of how LLMs work.
Each pattern is a specific claim: that a particular kind of interface move — a ghost text suggestion, an autonomy dial, an action log — solves a problem that a chat thread cannot. The pattern pages carry the side-by-side explanation and, where available, the live demo itself.
Every pattern is evaluated against four dimensions: Directability (can you steer it?), Observability (can you see what it is doing?), Predictability (do you know what comes next?), and Explainability (do you understand why it decided that?). These are not features. They are interface obligations.
Source Material
Each pattern in this archive is a specific claim about UX. Some claims were suggested by shipping products; others were sharpened by research papers. The list below is the working bibliography — the papers that explicitly informed one or more patterns in this collection.
K. J. Kevin Feng, Kevin Pu, Matt Latzke, Tal August, Pao Siangliulue, Jonathan Bragg, Daniel S. Weld, Amy X. Zhang, Joseph Chee Chang
Allen Institute for AI + University of Washington · 2024 · refreshed 2026-02 · arXiv:2412.10999
Mixed-initiative framework where humans and AI flexibly share planning *and* execution. Co-planning beats both full automation and full manual on outcome quality and user trust.
Madeleine Grunde-McLaughlin, Hussein Mozannar, Maya Murad, Jingya Chen, Saleema Amershi, Adam Fourney
University of Washington + Microsoft Research · 2026 · arXiv:2602.16844
Argues for rhythm-driven over interrupt-driven oversight, and treats concurrency as a designed parameter rather than an emergent property. Source for two patterns in the archive.
Hussein Mozannar, Gagan Bansal, Cheng Tan, Adam Fourney, Victor Dibia, Jingya Chen et al.
Microsoft Research · 2025 · arXiv:2507.22358
Concrete system for human-in-the-loop agentic workflows: co-planning, per-action-class guardrails, intervention mechanisms. Practical UI patterns for agentic oversight.
K. J. Kevin Feng, David W. McDonald, Amy X. Zhang
University of Washington · 2025 · arXiv:2506.12469
Defines five discrete autonomy modes (Operator / Collaborator / Consultant / Approver / Observer) with corresponding human-side control modes. A named-modes vocabulary that complements continuous per-tool dials.
Nenad Tomašev, Matija Franklin, Simon Osindero
Google DeepMind · 2026 · arXiv:2602.11865
Framework for safe and scalable delegation: cognitive friction, authority gradient, verifiable completion, and task-bid markets as the missing UX layer for cross-org agent work.
Kuangzhe Xu, Yu Shen, Longjie Yan, Yinghui Ren
Cyberspace Security University of China · 2026 · arXiv:2603.21735
Independent argument for *scaffolded* friction designed to preserve epistemic sovereignty — same primitive as Cognitive Friction with a different load-bearing rationale (cognitive agency over operational safety).
Blessing Jerry, Lourdes Moreno, Paloma Martínez
Universidad Carlos III de Madrid · 2026 · arXiv:2602.13745
Argues for human oversight as a first-class design principle in generative AI interfaces — not an afterthought. Frames accessibility as part of the oversight surface, not a separate concern.
Yunhao Luo, Arthur Caetano, Avinash Ajit Nargund, Tobias Höllerer, Misha Sra
UC Santa Barbara · 2026 · arXiv:2602.01481
Empirical study: users value agency preservation but accept assistance when it improves outcomes. Nuanced data on mixed-initiative perception — calibration evidence for any control-aware UX argument.
Nahal Mafi, Sahar Maleki, Babak Rahimi Ardabili, Hamed Tabkhi
UNC Charlotte · 2026 · arXiv:2603.04552
UX for AI must transcend frontend usability to encompass backend performance, org workflows, and societal decision structures. The frame this archive's patterns operationalise at the interface level.
Adrien Baranes, Rob Marchant
Google DeepMind · 2026
The cursor becomes the deictic anchor; voice becomes the verb. A research post with experimental demos in Google AI Studio; the shipped spin-off, Googlebook's Magic Pointer, takes typed prompts.
Adi Leviim
UX Collective · 2026
AI products inherited the messaging-app skeleton — branching, but no Cmd+Z. Article argues the 1987 undo primitive is the missing affordance and shows what an output-level Cmd+Z would look like.
Don Norman
Basic Books · revised 2013 edition · 1988
The feedback and visibility principles every interface inherits. Output Undo, Action Audit, and most Observability patterns are direct applications of "the system must reflect its state back to the user."
Ben Shneiderman
IEEE Computer · Vol. 16, Iss. 8 · 1983
The original case for visible, continuous, reversible direct manipulation over command-line abstraction. The patterns that move from prose-prompt to controllable surface are this argument applied to LLM I/O.
Bret Victor
worrydream.com · long-form essay · 2006
Distinguishes information software (you read) from manipulation software (you operate) and argues most products misclassify themselves. The case for generated answers that are interfaces, not paragraphs.
Marc Levoy
Stanford University · Proc. SIGGRAPH '94, pp. 139–146 · 1994
The grid recast as a substrate for expensive computation rather than arithmetic — cells hold images and volumes, so "larger formulas, longer computation times, and more complicated intercell dependencies" become the design problem. Answers it with the two extensions the fan-out pattern still runs on: cells can be selectively disabled, and multiple cells edited at once.
Cynthia Breazeal
MIT Media Lab · Robotics and Autonomous Systems, Vol. 42, Iss. 3–4, pp. 167–175 · 2003
Names the para-linguistic envelope display — gaze, brow, lean — as the channel that hands the speaking floor back and forth, and shows it works on timing rather than content: the same silent pause reads as “still mine” or “yours now” depending only on where the robot looks. Kismet held about 82.5% clean turns with four naive subjects and no shared transcript.
Guy Hoffman, Cynthia Breazeal
MIT Media Lab · IEEE Transactions on Robotics, Vol. 23, Iss. 5, pp. 952–961 · 2007
Separates perceived fluency from task efficiency: with outcomes statistically indistinguishable, subjects rated the anticipatory robot far higher on commitment to the team (5.0 vs 2.8, p<0.005), and the only measurable differences were temporal. Grounds the clock, not the channel — cite it for timing being what gets read, never for sub-second thresholds.
Google, now Agentic AI Foundation (Linux Foundation)
Industry standard · 2025
Open spec for inter-agent communication with optionally signed Agent Cards (JWS). It signs who an agent is, not what it delivered; task-outcome attestations are still open community proposals.
Anthropic
Anthropic · product announcement + research note · 2024
First production model trained to operate a computer interface directly. Turns every UI question (preview, audit, gates, friction) into a concrete shipping setting and provides the destructive-action-confirmation primitive several patterns build on.
Maggie Appleton
maggieappleton.com · talk + essay · 2024
Argues for structured surfaces that contain the "squish" of probabilistic generation. The practitioner-side companion to the academic patterns in this archive.
Geoffrey Litt, Paul Sonnentag, Peter van Hardenberg, Alexander Obenauer
Ink & Switch · 2024
Argues software should be end-user-modifiable as a default, with LLMs as the editing surface. The frame behind Generative UI, Smart Document, and the broader post-chat document model.
Adam Wiggins
Elicit (Ought) · engineering blog · 2024
First-party account of the fan-out: papers in the rows, extractions in the columns, and "fifty papers with twenty data extractions is already 1,000 individual cells to be generated". Names both halves the pattern needs — click a cell to flip through the quotes behind it, and treat per-cell errors and retries as ordinary rather than exceptional.
Andrej Karpathy
Essay (2017) + Y Combinator talk (2025) · 2017 · refreshed 2025
Frames neural-network weights as a new programming substrate (2.0); the 2025 follow-up extends that to "Software 3.0" — prompting as the new programming layer. The meta-frame for why interface design is now the bottleneck.
Saleema Amershi et al.
Microsoft Research · 2019 · refreshed 2026-07
The canonical 18-guideline taxonomy for human-AI interaction. Strongest anchors: G9 efficient correction, G11 why-did-it-do-that, G17 global controls. Its three zero-coverage guidelines — G6 bias mitigation, G14 cautious updates, G18 change notification — became this library's Bias Flag and Model Changelog patterns in July 2026. Nearly the whole Agent Workflow category sits outside its 2019 single-assistant frame.
Google PAIR team
Google Research · 2021 · refreshed 2026-07
Six-chapter first-party guide. Its Explainability + Trust and Feedback + Control chapters map almost one-to-one onto our Trust & Verification and control patterns; its confidence-display guidance ("determine if, when, and how to show confidence") is the softer default our confidence-signals avoidWhen encodes.
Sangho Suh, Bryan Min, Srishti Palani, Haijun Xia
University of California, San Diego · UIST '23 · 2023 · arXiv:2305.11483
Names and evaluates semantic zoom as an LLM interaction: the response is fed back to the model and re-rendered at three levels — Lines / Summary / Keywords — rather than cropped, with depth either bound to canvas zoom or pinned per node. The empirical case that a single depth axis over one answer is usable, not merely appealing (12.58 uses per session, SD 8.92). Pairs with the NNGroup study below: that one names the failure mode, this one designs and measures the fix.
Sarah Gibbons, Tarun Mugunthan, Jakob Nielsen
Nielsen Norman Group · 2023
Empirical usability study naming the exact chat failure modes — expanding/condensing whole answers when only a span needs work, manually harvesting fragments across regenerations — that selection-scoped transforms and single-axis depth controls exist to fix.
Raluca Budiu, Feifei Liu, Emma Cionca, Amy Zhang
Nielsen Norman Group · 2023
Diary study of real chatbot use: only ~4% of turns include a clarifying question and only ~22% of AI answers get fact-checked — the empirical backbone for asking-before-guessing and for making verification cheap enough to actually happen.
Caleb Chan
Nielsen Norman Group · 2025
Recommends against exposing step-by-step reasoning as a trust mechanism: displayed traces are typically post-hoc rationalizations, not the model's faithful process — lean on sources and stated limitations instead. Directly sharpened reasoning-trace's risks framing.
This is the working bibliography behind the archive. A source lands here when it sharpened a specific pattern claim — not for completeness, and not as a citation contest.
Why It Exists
The systems being built now can browse, plan, reason, and act. The interfaces around them still expect you to express everything as a prompt and receive everything as a response. That asymmetry is the design problem this archive is focused on.
The intent is to make the alternative concrete: not to describe better AI UX in the abstract, but to articulate and document specific interaction moves — patterns that transfer across product contexts because they address structural problems, not product-specific ones.
The research section exists to hold the longer argument around those moves: notes, essays, references, and framing that help explain why a post-chat interface vocabulary is needed in the first place.
Publishing Tracks
Browse WritingSmaller observations, references, and framing pieces that do not need a full essay format.
Longer writing on control-aware UX, post-chat interaction models, and agent interface design.
Quotes, examples, and structural comparisons that support the broader argument without turning the research section into a demo index.
A hosted newsletter link is configured and can be used without adding custom backend infrastructure.
Pragmatic Next Step
The next meaningful gain is density: better articulation, more essays, more notes, and a clearer archive of interface thinking. Backend complexity can wait until the site has something substantial to support.