Designing for AI
A new core UX skill: designing the loop
From Andreia Mesquita:
"Traditional UX had a clean contract: a human wants something, you design the path, and the human does it. The interface was a translator between intention and action. That contract is dissolving."
With agents, AI systems are moving from assisting to acting. In this case, the most consequential design decision isn't about the interface: it's about who stays in control, and when. Andreia lays out four positions on the human-AI control spectrum:
Human-in-the-Loop (HITL), where a human approves every consequential action
Human-on-the-Loop (HOTL), where AI acts autonomously but a human monitors and can intervene
Human-AI-in-the-Loop (HAITL), where the two genuinely co-create in real time
Out-of-the-Loop (OOTL), where the AI runs solo
Each position comes with its own potential design trap: HITL breeds approval fatigue, HOTL invites automation complacency, HAITL blurs authorship, and OOTL demands rigorous audit trails even when nobody is watching. A key factor is reversibility: the less reversible the outcome, the more human involvement you need. With agents, we're designing trust calibration, handoff points, escalation triggers, and accountability chains.
Designing for AI
Six possible futures for apps in the age of AI
In a recent LinkedIn post, Dan Saffer explores six ways apps can evolve in an AI world:
Some apps vanish entirely because they provide a function that AI can fully absorb.
Some become invisible infrastructure, feeding data to AI platforms through APIs while their interfaces disappear.
Some survive as trust marks: proof that something is real and licensed, not generated.
Some shrink to "contextual slivers" that appear and recede without you ever opening them (for example, your workout playlist queued when you walk into the gym).
Some persist because people stay for other people, not for the tool (social networks).
And some keep doing exactly what they do now, just with AI layered in.
Dan maintains it's not which future wins: it's which configurations your product can actually be great at, and you probably only get to pick two or three.
Dan Saffer | LinkedIn post
⚡ Quick Read (2 minutes)
Building with AI
Figma and the new competitive landscape
Richard Best argues that Figma is showing the same kind of incumbent complacency that once left Adobe exposed to Figma: dominant market share, slower product velocity, and too much reliance on ecosystem workarounds. In a sharply argued essay, he points to long-requested functionality like the recently-released Slots, plugin dependence for important workflow needs, and unresolved rough edges in Figma’s MCP design-to-code story.
The piece is also useful as a map of what’s emerging around Figma:
Pencil stores design files as open JSON-based .pen files inside your repo.
Google Stitch is quickly expanding from UI generation into a more capable AI-native canvas workflow.
Paper is building around code-connected design and native Tailwind export.
The throughline is that newer tools are increasingly (and natively) treating design and code as parts of the same system, rather than separate artifacts connected by handoff.
Richard Best | Figma is burning its moat one token at a time
☕ Medium Read (10 minutes)
AI Agents
Building a context infrastructure with Claude Cowork
Eric Porres spent two months using Claude Cowork as his primary working environment: not as a chatbot, but as a full workstation. His core argument is that the bottleneck with AI agents is the absence of structured context, including preferences, routing rules, and institutional knowledge that turn AI into a useful collaborator.
The piece lays out a five-layer architecture: MCP connectors, a CLAUDE.md file, skills, reference data, and plugins. Each layer builds on the ones below it. His email triage works better because a family assistant skill gives context about his kids' schools. His newsletter drafts improve because a voice guide encodes his writing cadence. The power isn't in any single component; it's in how the layers work together.
Eric provides both the framework and working examples, particularly helpful if you've not yet created your own agents.
Eric Porres | Claude Cowork might be the most consequential piece of software you're not using
☕ Medium Read (12 minutes)
For more on skills, check out Nate Jones on What's really happening inside the skills ecosystem (watch time 26 minutes)
That’s it for this week.
Thanks for reading, and see you next Wednesday with more curated AI/UX news and insights. 👋
All the best, Heidi
