Skip to content
Work About Publications Talks Contact
Beyond Chat.

Make the invisible work visible.

AI can write, plan and act. The interface still asks us to squeeze intent, context and control into one text box. I built a pattern library to explore what happens when that work gets its own surface.

Explore the case Open the live library
Role
Research, framework, design and code
Timeframe
2024–ongoing
Context
Independent research · 80 patterns in five families
Status
Public library · research ongoing

01 / Starting point

A single field is doing six jobs.

For a quick question, chat is useful. For longer work, one field has to carry intent, constraints, context, feedback, control and revision. People must remember what the system knows, what it will do next and where they can intervene.

Claude interface with a central chat input AI app builder interface with a chat input Perplexity interface with a chat input ChatGPT interface with a chat input Manus interface with a chat input DeepSeek interface with a chat input
Six tools, one pattern: a blinking cursor in a box. Claude, an app builder, Perplexity, ChatGPT, Manus and DeepSeek, as shown in the original talk. · Actual interface

The design question became practical: which parts of AI work should be visible, steerable and checkable before they disappear into a prompt?

02 / Direct control

Change the work without rewriting the request.

The Property Panel turns part of a prompt into direct controls. This small illustration lets you change a sample response; the full prototype ↗ explores the pattern in more depth.

Draft a project update

“Summarise the status for the team. Keep the facts and dates intact.”

The project is on schedule. The team confirmed the next review for 12 October. Two decisions remain open: the final content order and the handover date.

Example source: fictional status note

The content stays explicit. Form, detail and evidence become decisions you can see and revise. · Illustration

03 / Framework

Control starts before the interface.

The research separates a principle from the process and the controls that carry it. That distinction matters when AI moves from answering questions to acting on someone’s behalf.

  1. Why. Keep judgment with people.

    AI can open possibilities. People should retain the authority to select, intervene and take responsibility for what is done.

  2. How. Scaffold intent and delegation.

    Structured choices and adjustable autonomy make the scope of a task clear before and during the work.

  3. What. Make control operable.

    Visible state, scoped actions and evidence give that principle a concrete place in the product.

Whatinterface Howprocess Whynormative
Three layers of control: normative, process and interface. The model appears in my writing and talks: Bitkom essay → · talk → · Illustration

Automation is a decision at every stage.

A trip planner can do substantial work before anything is booked. Each stage shows where the interface can prepare, reveal and hand a decision back to a person.

Easy to reviseApproval before commitment

  1. Acquire

    Gather routes and charging stops.

    AI prepares · person can inspect
    Interface move: visible sources and freshness

  2. Analyse

    Make the trade-offs legible.

    AI compares · person can adjust
    Interface move: comparable options and editable constraints

  3. Decide

    Select an itinerary.

    Person chooses · AI proposes
    Interface move: editable plan and explicit choice

  4. Execute

    Confirm before booking.

    Person approves · system acts
    Interface move: review and confirm

Illustrative trip-planning scenario from the BMW UXCOMM talk. The degree of autonomy depends on the risk and reversibility of each action. · Illustration

Four properties. One overlap.

The 2025 essay asked whether an AI system could be observed, anticipated and directed. As the pattern library developed, explainability became a fourth question. DOPE is a design lens for the interface layer.

The DOPE framework Four overlapping circles for Directability, Observability, Predictability and Explainability. Their shared centre represents trustworthy AI autonomy as a design goal. TRUSTWORTHY AI AUTONOMY AI

  1. Directability

    Can I steer the work? Give people a way to edit the plan, set boundaries and change course while the system is working.

  2. Observability

    Can I see what is happening? Show the current state, the sources in use and what the system has already done.

  3. Predictability

    Do I know what comes next? Preview the next action and its consequences before the system commits to it.

  4. Explainability

    Can I check the answer? Connect an output to evidence and be clear about what the explanation can and cannot establish.

Trustworthy AI autonomy is the ambition at the centre, not a measured trust score. The full DOPE framework ↗ is in the live library. · Illustration

04 / Pattern library

From a principle to a vocabulary of interfaces.

Beyond Chat collects interaction patterns already appearing in products alongside emerging and speculative directions. Every entry pairs a concrete idea with a prototype, its use case, risks and examples.

Each prototype sets the pattern against a baseline on the same task, usually the chat default: the demo is the comparison. Here are four of them, one for each property.

  1. Editable Plan

    Before the agent runs, its plan is a document you can edit. In chat the plan is negotiated in prose; here a step is removed, another moves up, and the agent runs only what was approved. Open the prototype ↗

    Directability: can I steer the work?

    Editable Plan in the live library, October 2026; on phones the pattern side only. · Actual interface

  2. Context Gauge

    Chat shows an infinite scrollback while the model sees a sliding window. The gauge shows how full the window is and which sources occupy it, warns before compaction and logs what has left. Open the prototype ↗

    Observability: can I see what is happening?

    Context Gauge in the live library, October 2026; on phones the pattern side only. · Actual interface

  3. Intent Preview

    Before an agent modifies code, it shows every planned step with risk levels, affected files and scope. Here the high-risk step is skipped before a single line changes; in chat, people find out after the fact. Open the prototype ↗

    Predictability: do I know what comes next?

    Intent Preview in the live library, October 2026; on phones the pattern side only. · Actual interface

  4. Citation Trail

    Hover any sentence in an answer and the contributing sources highlight in place. Chat leaves a pile of footnotes; the trail binds each claim to where it came from. Open the prototype ↗

    Explainability: can I check the answer?

    Citation Trail in the live library, October 2026; on phones the pattern side only. · Actual interface

Five families across the AI workflow, 80 patterns in the catalogue, with one preview tile from each family. The library is designed to be explored: each prototype is a starting point for a design decision, not a universal recipe. Browse all patterns ↗ · Actual interface

05 / Research & iteration

Research changed the design.

I began with writing about autonomy in design, then studied seven designers’ AI-assisted work. The public library turned those questions into prototypes. Subsequent research and product-example audits exposed claims that needed to become more precise.

  1. Frame the question

    Essays and a qualitative study with seven professional designers explored curation, autonomy and responsibility in AI-supported design.

  2. Build the vocabulary

    Five pattern families, a shared schema and interactive demos made the ideas comparable across writing, visual work, agents, trust and memory.

  3. Audit the claims

    Research reviews and a targeted audit of example rows checked whether references actually demonstrated the interaction described.

Today’s Intent Preview in the live library: a light interface with the same task and the same six steps
The early, dark version of Intent Preview with the same task and steps
Intent Preview, early and today: the same task and the same six steps, as a dark prototype on the left and in the live library on the right. Drag the handle to compare. · Actual interface
Confidence Signals, early version: in chat every claim sounds equally certain, the pattern gives each claim a confidence level and its sources. · Actual interface
Generative UI, early version: in chat the answer is a block of text, beside it the same answer as key figures and a chart that changes from bars to lines to a ring. · Actual interface
Supervisor-Worker, early version: a single agent writes one long answer, beside it a supervisor splits the task among worker agents and shows each one’s progress. · Actual interface
  • mature
  • emerging
  • aspirational
Evidence status of the 80 patterns in the live catalogue, distinct from whether their catalogue pages are live. · Live data from beyondchat.design

Displayed trace

An account of a result that a person can inspect.

is not

Internal process

The route the model actually took.

The design response: treat traces as checkable claims and pair them with sources and limits. · Illustration

“Reasoning Trace” needed a narrower claim.

A displayed chain of reasoning can be a useful account of a result, but it is not necessarily a faithful window into how a model arrived there. The pattern’s risk language was revised after the research review.

That correction matters: a pattern library should help teams ask better questions, including where its own examples are uncertain.

Read the research notes ↗

06 / What remains

A working library, and a question worth testing.

Live at beyondchat.design with 80 catalogue entries in five families, an essay and a talk. A study platform that compares chat with structured controls is built; testing is still ahead.

Reflection

I built the vocabulary before I tested it. If I started again, I would run a small comparison study alongside the first ten patterns, not after the eightieth.

A mosaic of Beyond Chat interaction pattern previews
Selected pattern previews from the library. · Actual interface

Next project

Plooi

A chat forgets, a canvas remembers — a collaborative node editor that turns design methods into AI workflows.

View project →