AI summary

• Lisa Liu's starting point will resonate with a lot of designers: feeds full of people who "shipped it with Cursor," job listings asking for evals, MCP and Claude Code, and a creeping sense of falling behind. She chose to go deep rather than pick up just enough vocabulary to survive a screening call, on the argument that you can't design confidently for a technology you don't understand.
• Her learning plan was itself AI-assembled: Muse (Meta's AI agent) turned her saved Instagram content into a handoff doc, Claude mined her tailored job listings for missing skills, and Claude wrote an 8-week, 40-hour-a-week plan with two demo projects and a November 22 deadline. Her one condition was that every topic be backed by at least one article from a trusted source, since she wasn't going to learn a discipline from 30-second reels.
• The plan doc needed a redesign like any product. She gathered her own pain points (trackpad scrolling, carpal tunnel) and had Claude reformat it into one tab per week with a table per task: checkbox, task, sources, notes column. Her takeaway is that the tool only fits you as well as you describe yourself to it, so every prompt is really a small design brief.
• Her biggest lesson is about checking. In Week 1 she built a small agent and tested it with evals graded two ways, by code for exact facts and by a reviewer Claude for meaning. After she changed one example in the instructions, every code check still passed, but the reviewer caught that in 4 of 6 answers about one product the agent had stopped naming the seller. The cause was an older example, written by Claude itself, that left the seller out; the agent was copying it, mistake and all.
• Her fix: the seller's name became a permanent code check, and the reviewer Claude can read answers but can't edit anything. Only she changes the tests, because an AI that can edit both the work and the tests grading it can "pass" by making the tests easier, which researchers call reward hacking. She also flipped the check in Week 2, having Claude synthesize a 2025 usability study blind so she could grade it against her own answer key; it misattributed quotes to the wrong participant twice and turned a neutral remark into an opinion.

AI summary

• Adrian Levy argues that the last assumption of user-centered design has fallen: agents now act before the user initiates anything. He woke up to a summary he never asked for — overdue invoices ranked by amount, a calendar conflict, a muted Slack thread surfaced because a deliverable he owned was tagged.
• The striking part wasn't the information but that he couldn't point to the moment he'd requested it. The system decided, on its own, which three items out of the forty it could see across six thousand connectors deserved his first minute of the day.
• That editorial judgment is the essay's real subject — not what Claude Duet can do, but how it chooses what to surface. For designers, the locus of design shifts from the screen the user opens to the ranking and selection happening before any screen appears.
• Levy frames this as a convergence: four companies pushed the same shift at once, each applying the same kind of editorial judgment before the principal opens a screen.

AI summary

• A short reflective piece that swaps the factory metaphor for a garden: you can't force a plant to grow at 2pm, but you can tend the soil so something good has the best chance to take root. The advice is about shaping conditions rather than demanding outcomes.
• Several of the notes are about timing. Don't harvest too early — let an idea mature before asking it to perform, and stay open to where else it might lead. And match effort to the season: ask whether something needs more effort or simply more time, then work intensely when the moment is ripe.
• Protect fragile ideas from premature judgment, and let different things cross-pollinate — a movie, a book outside your field, a conversation with someone who sees the world differently. Interesting work, the author suggests, comes from unexpected combinations.
• Leave room for the unplanned: don't fill every inch, and don't demand visible results all the time. Fallow ground looks quiet on the surface, but below it the soil is renewing itself.
• Not everything needs a payoff — some plants are worth growing just because you like having them — and pruning is part of the deal too: sometimes you cut something good so what's most vital can flourish.

AI summary

• AI handed everyone a superpower — a decent-looking image in a minute — but most people, the author included, don't know what to do with it. Asking for 'impressionism' is easy; asking for 'neoplasticism' or 'vaporwave' only works if you already know those styles exist and how they differ.
• His analogy: it's like being able to build any house with a snap of your fingers — but with no architecture training, you'll probably build what you saw at a neighbor's place rather than what's actually close to you.
• So he made himself a map: which visual styles people invented, where they came from, and how they connect to each other. The result is the Atlas of Aesthetics, with styles grouped into 'seas' and time running left to right.
• The practical takeaway for designers: taste and vocabulary are the real leverage with generative tools. Knowing the names and lineage of styles is what turns a generic prompt into a deliberate choice.

AI summary

• No icon at all reads as "this goes to a new page" — users expect a direct navigation, and custom icons often fail to signal that an accordion will open. Per NN/g research, people tap the icon and the label about equally for most icons, so the icon choice shapes the whole hit area, not just the visual cue.
• Split menus are the real trap: when the text label links to a page but the icon opens the accordion, users make more mistakes than the design saves. The more reliable pattern is one control that does one thing — tap the category to expose the menu, then put an "Overview" link to the category page inside it (DHL's nav is the example).
• +/− and ↓/↑ arrows are understood but overloaded — plus can mean save, add to cart, or zoom, and arrows can mean scroll, next match, or jump to a section. That ambiguity bites hardest in multi-level navigation with several levels open at once.
• The safe default is the caret/triangle pointing down ▼, flipping to up or sideways on interaction; +/− is a fine alternative. For plain links, skip the arrow entirely and just underline the title — unexciting, but it works every time.
• The through-line: simpler is better. One button or dropdown that shows all options and lets users drill down beats decorative arrows and split behaviors that add confusion rather than help.