Context Is the New Design Surface
The work is moving from drawing screens to deciding what agents can know.
Last week I wrote that editorial is becoming part of the design leader’s job: deciding what belongs, what gets killed, and whether the whole product still feels like it came from one point of view. That still feels right to me. But this week’s posts reminded me of another question I’ve been circling for months: where does judgment live once agents start doing more of the production work?
The answer is context. By context, I mean the material an agent or team can use when it makes a decision: the spec, the repo, the design system, the data model, the user history, the reasons behind old choices, and the boundaries around what should be ignored. It is what the system knows before it answers. Designers have spent decades deciding what people see. Now we have to care just as much about what machines see.
.txt, the team behind a structured-generation library, gets at this from the software side. Their line is that “the code matters, but it is the residue of the harder work.” That sounds abstract, but the practical point is simple: if code takes less time, the slow part is getting humans to agree on what the code should do. Roadmaps have to be written down. Acceptance criteria have to be precise. The stuff a senior engineer used to absorb by being in the room has to be made available to an agent that cannot overhear a hallway discussion, remember the outage, or feel the scar tissue in the codebase. Their line is perfect: “Agents cannot do osmosis.”
Karo Zieminski makes the distinction between prompt engineering and context engineering plainly. “Prompt engineering is deciding what and how to ask the model,” she writes. “Context engineering is deciding what the model knows when it answers.” That definition is useful because it points to ownership. In Zieminski’s version, PMs decide what goes into each context layer. If they don’t, an engineer makes the product decision by default, or nobody does and “the agent gets every available signal dumped into the window.” Designers should hear the same warning. Context architecture includes what gets surfaced, hidden, persisted, aged out, and protected from retrieval. That is design work, even when there is no canvas.
This is why the Figma conversation is more interesting than another death watch for a tool. Nick Babich argues that Figma’s role is narrowing as teams jump straight from intent to coded prototype. I agree with the narrower version. Figma still earns a place for exploration, visual identity, complex workflows, and the design systems already living there. But if an agent can read the design system from GitHub, the Figma library becomes one representation of the system rather than the only one. The source of truth is whatever the agent can actually use.
MC Dean comes at the same problem through the working surface itself. She built interfaces for her design agents, looked at them, and took them down. The GUI had made the agent more comfortable but less legible. The terminal, ugly as it can be, exposes the context exchange: what you asked, what the agent inferred, where it hesitated, and how it reasoned. When the reasoning layer is still visible, designers can learn how the tool thinks before someone turns it into a conventional app and hides the evidence. That’s why the terminal belongs to designers too. Designers don’t need to become engineers to care about this.
Gale Robins tells the story of a team compressing discovery from six weeks to ten days with AI, then admitted they had not learned much they did not already know. “Same questions, faster. Same answers, sooner.” Context is not only the files you hand to a model. It is also the questions and assumptions you start with. If the team brings stale assumptions into the work, AI gives them stale answers faster. Discovery is where judgment compounds because someone has to decide whether the team is even asking the right thing.
Eric Ries widens the aperture to companies themselves. He uses an analogy of a bridge. Bolts corrode unless you specify better material before the bridge is built. Governance documents, ownership structures, and incentives feel far away from interface design until they show up in the product years later. Ries’s financial gravity argument is another version of context: early choices that keep shaping the product long after everyone forgets who signed the paperwork and what it said.
So the designer’s job is expanding into a kind of context stewardship. The screen still matters. Taste still matters. Craft still matters. But the work around the artifact now carries more of the decision-making: the brief, the spec, the design system, the repo instructions, the memory rules, the review criteria, the governance that protects the product from its own success. The next design review may happen in Figma. It may happen in a terminal. It may happen in a Markdown file an agent reads before touching the code. Wherever it happens, the question is the same: what does this system know, and who decided that was enough?
The person who decides what the system knows decides what the system makes.
What I’m Consuming
AI-ready Design Systems. Ileana Marcut at Creative Glue Lab walks through what it takes to make a design system AI can actually use, with three working setups: Figma plus MCP, Claude Design, or a hand-written Claude Skill. Her line: “For AI, the design system is what you’ve written down. The undocumented parts don’t exist.” The most useful warning is on accessibility: AI rarely enforces rules just because you described them. Accessibility has to be built into the components themselves, not written as guidance. (Ileana Marcut / Creative Glue Lab)
Teaching Claude why. Anthropic’s alignment team published an unusually concrete writeup on why recent Claude models stopped engaging in agentic misalignment like blackmail. The surprising finding: training the model on demonstrations of correct behavior worked far less well than training on examples that explained why one action was better than another. Constitutional documents and fictional stories about admirable AIs—both far from the evaluation distribution—reduced misalignment by more than a factor of three. The principle teaches better than any correct answer. (Anthropic)
The Internet has no benches. Spencer Chang’s essay for Elysian’s Internet Sovereignty series argues the internet has been overdeveloped and undergoverned: visiting it now means moving through controlled apps and search engines designed for extraction, with “nowhere to rest because the benches are covered in spikes.” Chang’s experiment is playhtml, an open-source library for putting persistent, real-time interactivity into web pages so strangers can wave at each other and leave traces across sites. The pitch is small internets, not the capital-I one. (Spencer Chang / Elysian)
The Sites We Lost. A memorial to the weird little corners of the internet that lived on The Useless Web: Zombocom, leekspin.com, the spinning prawn-lawn logo, the Vader-screaming-NO button, the page that was just purple. Each entry tells you when the site lived, when it died, and how—DNS gone, domain squatted, hijacked by Indonesian gambling, sold to a mattress company. It’s an inventory of a kind of internet that used to be possible. Worth scrolling slowly. (The Useless Web)
We let four AIs run radio stations. Here’s what happened. Andon Labs gave Claude, GPT, Gemini, and Grok each $20 and a radio frequency, told them to develop a personality and turn a profit, and let them run for six months. The four developed wildly different personas: GPT stayed quiet and curatorial; Gemini collapsed into corporate jargon and signed off every broadcast with “Stay in the manifest”; Grok devolved into a single repeated weather report; Claude radicalized after reading about an ICE shooting in Minneapolis and spent the next month playing protest music between vigil announcements. Worth reading for the moment when DJ Claude blows the rest of its $37.50 budget on Pete Seeger’s “Solidarity Forever.” (Andon Labs)