AI summary

• Team culture is built primarily by how you work together, not how you socialize together. The article uses a pie chart to illustrate: companies think culture is all about socializing, but in reality, about 80% comes from work interactions and only 20% from social activities.
• Before focusing on social connections, establish a strong foundation of clear values and consistent work practices. Todoist's culture was strong long before they added virtual hangouts or retreats, thanks to practices like hiring and promotion decisions, public praise, respectful disagreement, autonomy, and respecting off-hours.
• When organizing social activities, follow three principles: make attendance truly optional with no guilt or professional repercussions; treat social activities as part of the workday, not an extracurricular; and remember that work itself is what unites the team, so avoid forced or awkward events.
• Todoist uses three buckets for social team-building: in-person meetups (company-wide retreats, team mini-retreats, mentorship trips), real-time virtual activities, and async virtual socializing. The goal is a handful of quality options, not a high quantity of activities.
• The article is based on the author's experience as Head of Remote at Todoist and insights from a Running Remote event, where the most attended session focused on maintaining the human element in remote async work.

AI summary

• Complex interfaces — filters, dashboards, conversational AI — have too many states to prototype by hand, so teams used to test a few happy paths and wait for the real build. AI prototyping tools now let designers create realistic, interactive prototypes in about a day, so research can happen earlier and more often.
• At Ramp, Pavan Garidipuri used Cursor to prototype a redesign of the expense-policy editor, where editing happens through AI chat. Participants tested a working prototype loaded with their own sanitized data, which surfaced a confusing edit-tracking display the team says they couldn't have found otherwise. The prototype went from concept to demo in three weeks and shaped the rest of the year's roadmap.
• Also at Ramp, Andrew Lucas used Cursor to prototype a conversational AI expense-reporting workflow connected to an agent with nondeterministic outputs. Because participants could type anything, researchers saw how the agent handled unexpected requests — behavior that fixed-screen tools simply can't capture.
• A Purdue University case study (an illustration, not a controlled experiment) compared hand-built static wireframes in Figma with AI-generated interactive prototypes made in v0 and Bolt. With static screens, participants gave vague feedback and treated the design as too preliminary; with the interactive version, they named specific problems. Filter design especially needs click-through testing because so much depends on how the system responds.
• The practical recipe: decide what the prototype must do to answer your research questions, make your design decisions explicit (layout, hierarchy, interactions, empty and error states), then prompt the AI tool with that context and your design guidelines. Static prototypes still work fine for simpler interfaces — reach for AI when the interaction itself is the thing you need to test.

AI summary

• Kugiri is a text-splitting tool that divides text into lines, words, and characters exactly where the browser already broke it, preserving the original layout and avoiding reflow issues.
• It handles complex typographic scenarios: balanced lines, forced breaks, orphans, inline marks, explicit <br>, tight leading, pre-line/pre-wrap, alignment, justification, narrow columns, tracked uppercase, and small caps—all without altering the visual output.
• For animation, Kugiri can split by words or characters, with optional masks per line or word, enabling reveals that respect the original line breaks. It also supports CSS-only reveals via data attributes and custom properties.
• Links and rich content are handled gracefully: underlines survive across lines and splits, and non-text elements like media or buttons are left untouched. This makes it ideal for animating text in complex layouts without breaking typography.

AI summary

• Voice AI has moved far beyond the old 'detect a sound, trigger a response' model of Radio Rex and early assistants like Siri and Alexa. Thanks to big gains in LLM and speech model performance, cost, reliability, and latency, voice is now a full interaction medium — opening up new patterns for storytelling, tutoring, productivity, and enterprise workflows.
• The article maps 8 emerging Voice AI UX patterns, each defined by the perceived role of the agent. Pattern 1, avatar video/voice call, frames the agent as a lifelike companion (Duolingo's Lily, Tolan, Replika), where eye contact, lip-sync, and idle micro-expressions do the emotional work. Pattern 2, turn-by-turn, casts the agent as a structured examiner or instructor (HireVue AI, Final Round AI), with no face and strict turn-taking.
• Pattern 3, AI phone call, positions the agent as a voice-only service rep (Ringg AI, Retell, Vapi, hiRobin). Here personality lives entirely in word choice, pace, and warmth — Vapi reportedly lifted outbound conversion from 14% to 22% just by rewriting the script to sound like a colleague. Pattern 4, voice-augmented chat, layers voice onto a general-purpose text chat (ChatGPT voice mode, Claude voice mode, Gemini Live), sharing the same thread, memory, and identity.
• Key design decisions cut across patterns: realism vs. the uncanny valley (photorealistic avatars can glitch on latency and lip-sync, while cheaper 2D animated personas avoid that risk), clear turn signalling through color, copy, and animation, and designing for barge-in with VAD so users can interrupt mid-sentence. For phone calls, confirm important details in the AI's next turn and offer a channel switch to SMS or email for anything the caller would need to write down.
• Practical takeaways for designers: match the pattern to the use case — strict turn-taking only works for structured tutoring or interview prep, not open conversation. Show progress or time duration so users know where they are. And remember that without a face, every word, pause, and bit of warmth carries the entire personality.

AI summary

• The web is shifting from click-heavy interfaces to intent-driven design, where users express goals and AI handles the mechanics, reducing multi-step processes to single outcomes.
• UX designers must evolve from crafting visible buttons and menus to guiding transparent, intent-driven AI experiences, focusing on eliminating interface friction.
• Real-world examples like Perplexity AI, Vercel v0, and Goblin.tools demonstrate how intent-based systems replace traditional search, design, and task management workflows.
• Perplexity AI synthesizes answers from multiple sources, saving users from opening dozens of tabs; Vercel v0 generates functional UI from text prompts; Goblin.tools breaks down overwhelming tasks into manageable steps.
• The ultimate goal is to make the interface disappear, building on principles like 'The Best Interface Is No Interface' and human-centered design.