# Market research: UI patterns for personal AI daily-work assistants (2025-2026)

Scope: single-user AI life managers - daily planning, journaling, habit/health tracking, voice capture, proactive briefs, chat + structured views. Feeds the Clawsy redesign.

## Messaging-first proactive assistants

### Poke (Interaction Co)
- Pattern: the assistant lives inside iMessage/WhatsApp/SMS - no separate app. It reads connected email/calendar/files and turns messages into actions (draft replies, reschedule, pay bills), with typing indicators and one-tap replies to mimic texting a human. [TechCrunch](https://techcrunch.com/2026/04/08/poke-makes-ai-agents-as-easy-as-sending-a-text/), [Product Hunt review](https://www.producthunt.com/p/poke-by-interaction-co/a-week-with-poke-review-a-promising-start-for-a-proactive-ai-assistant)
- Works well: calendar scheduling was reviewers' most reliable single feature; the zero-app-switch, "feels like texting a friend" loop drives real daily use.
- Fails: breaks down on 3+ step chained tasks; weak at knowing what's worth telling the user proactively and at holding onto prior decisions; basic time-zone/day-of-week reasoning errors caused a late meeting reminder. One user: "still feels like working with a clever, caffeinated toddler sometimes." [Slashdot](https://slashdot.org/software/p/Poke/), [saner.ai review](https://blog.saner.ai/poke-reviews/)
- Steal: text-message affordances (typing indicator, one-tap confirm, short turns) lower the perceived cost of talking to an agent far below a "chat app".

### Dot (New Computer) - shut down Oct 2025
- Pattern: a single long-running, memory-accumulating chat companion - "a living journal that talks back" - with no structured views at all. [Fast Company](https://www.fastcompany.com/91142350/dot-an-ai-companion-app-designed-by-an-apple-alum-launches-in-the-app-store)
- Works well: long-horizon memory produced a felt sense of being known over time. [TechCrunch](https://techcrunch.com/2025/09/05/personalized-ai-companion-app-dot-is-shutting-down/)
- Fails: shut down after ~13 months; users described "grieving" the companion. Memory-only companion chat with no task/calendar/data utility is emotionally sticky yet product-wise fragile. [Medium](https://medium.com/artificial-synapse-media/ai-confessional-app-dot-shuts-down-after-4-months-leaving-users-grieving-digital-companion-fad5aecb4066)
- Steal: a warning, not a pattern. Clawsy avoids this by pairing chat with structured surfaces.

## Proactive daily-brief / scheduled-action features from the frontier labs

### ChatGPT Pulse (OpenAI) - shut down June 2026, ~9 months after launch
- Pattern: async overnight job reads memory, chat history, connected Gmail/Calendar, and produces a scroll of tappable "cards" each morning, mobile-only, Pro-tier. [OpenAI](https://openai.com/index/introducing-chatgpt-pulse/), [Yahoo](https://tech.yahoo.com/ai/chatgpt/articles/openai-launches-chatgpt-pulse-deliver-092118051.html)
- Works well: the card format (tap to expand, thumbs up/down, spin into a new chat) is a clean surface for heterogeneous content in one feed.
- Fails, and this is the single most important negative data point in this survey: killed in June 2026. Root causes: (1) relevance decay - it kept surfacing topics the user had already resolved weeks earlier; (2) the best models hit only ~64% accuracy on whether proactive help was wanted, so roughly a third of cards felt like spam; (3) a feed of reading material with no action attached "becomes homework". [prowlo.com](https://prowlo.com/blog/chatgpt-pulse-shut-down), [Tom's Guide](https://www.tomsguide.com/ai/i-test-ai-for-a-living-and-this-is-hands-down-the-worst-new-ai-tool-of-2025)
- Steal (negative lesson): inferred proactive feeds fail. A morning brief should be built from things the user or an explicit routine asked to track, and every card needs an obvious next action.

### Gemini Scheduled Actions + Daily Brief (Google)
- Pattern: user-configured recurring prompts that Gemini runs on a schedule (up to 10 active), landing as a manageable list in a dedicated tab rather than an algorithmic feed. [Google blog](https://blog.google/products-and-platforms/products/gemini/scheduled-actions-gemini-app/), [Digital Citizen](https://www.digitalcitizen.life/gemini-scheduled-actions-how-to-automate-tasks-with-google-gemini/)
- Works well: explicit opt-in per routine avoids Pulse's relevance-decay problem. [testingcatalog.com](https://www.testingcatalog.com/google-tests-new-scheduled-actions-manager-tab-for-gemini/)
- Fails: templates are generic; a flat list of scheduled actions doesn't compose into one coherent "here's your day".
- Steal: separate the scheduling primitive (a named, editable, cancelable recurring job) from the content it produces; the manager-tab UI is a good reference for a routines-settings screen.

### Claude Routines (Anthropic)
- Pattern: a saved prompt + tools configuration triggered by schedule, API call, or webhook, running headless. [Anthropic blog](https://claude.com/blog/introducing-routines-in-claude-code), [InfoQ](https://www.infoq.com/news/2026/05/anthropic-routines-claude/), [Builder.io](https://www.builder.io/blog/claude-code-routines)
- Works well: routines are automations of the same loop the user already runs manually, so there is no new UI to learn.
- Fails: aimed at developer use; no personal brief surface ships with it.
- Steal: routines as triggers, decoupled from a fixed output format. The missing piece is a visible, editable "your routines" list.

## Daily-planning UIs

### Sunsama
- Pattern: a guided morning-planning ritual - pull tasks into one inbox, estimate time, drag onto a calendar, set the day's realistic goal, and a daily shutdown ritual. [Efficient App](https://efficient.app/apps/sunsama)
- Works well: visual overload warnings when you plan too much; calm minimalist kanban+calendar hybrid. [Calmevo](https://calmevo.com/sunsama-review/)
- Fails: no automation; weak on mobile. [Productive with Chris](https://productivewithchris.com/tools/sunsama/)
- Steal: the two-ritual bookend (morning plan / evening shutdown) as a fixed daily shape, with an explicit overload warning.

### Motion, Akiflow, Reclaim.ai
- Pattern: fully automated AI schedulers that place tasks and habits onto the calendar and re-arrange them (Motion, Reclaim) vs. a single task inbox the user drags in (Akiflow). [Morgen comparison](https://www.morgen.so/blog-posts/akiflow-vs-motion)
- Works well: Reclaim's auto-defense of focus time; Motion's "just feed it deadlines".
- Fails consistently: the AI silently moving blocks and overloading days without breaks is a recurring complaint; billing anger compounds trust problems; Reclaim has no native mobile app. [Morgen Reclaim alternatives](https://www.morgen.so/blog-posts/reclaim-ai-alternatives)
- Steal: any auto-move of a user's calendar needs a visible diff and undo, not silent mutation.

### Amie
- Pattern: award-winning calendar+todo app that pivoted through 2025 to be primarily a bot-free AI meeting notetaker. [usecarly.com](https://www.usecarly.com/blog/what-happened-to-amie/), [Productive with Chris](https://productivewithchris.com/app-reviews/amie-review-2025/)
- Steal: bot-free, on-device capture as the default for conversation capture; cautionary tale about chasing the AI feature of the moment.

## Notes apps with AI

### Reflect
- Pattern: minimalist backlink notes; AI reads across the note graph and transcribes voice straight into the daily note. [downloadchaos.com](https://downloadchaos.com/blog/reflect-notes-ai-features-note-taking-innovation), [XDA](https://www.xda-developers.com/started-using-ai-powered-tool-for-all-my-notes/)
- Steal: voice to transcript lands directly in the day's note, not a separate recordings list. Capture location defaults to "today".

### Mem (Mem 2.0)
- Pattern: "second brain that organizes itself" with an agentic layer and voice capture. [saner.ai](https://blog.saner.ai/mem-ai-reviews/), [fahimai.com](https://www.fahimai.com/mem-ai)
- Steal: voice mode as a first-class capture button next to text, not a separate app mode.

### Notion AI
- Pattern: AI with full read access to the workspace; Q&A answers pull from real docs and databases with links back. [aidiscoveries.io](https://aidiscoveries.io/notion-qa-review-2025-the-ai-tool-that-reads-your-entire-companys-brain-so-you-dont-have-to/), [theaipicker.com](https://theaipicker.com/notion-ai-review/)
- Steal: "ask a question, get an answer sourced from your own docs/tasks/journal, with links back to the source" is the shape a Clawsy chat-over-workspace should take.

### Capacities
- Pattern: object-based knowledge management - Books, People, Projects, Meetings, Ideas as typed objects. [fahimai.com](https://www.fahimai.com/capacities-review), [academichelp.net](https://academichelp.net/note-taking-apps/capacities-review.html)
- Fails: steep learning curve, weak mobile.
- Steal: typed objects as the retrieval unit, matching Clawsy's entity model.

### Tana
- Pattern: outliner + supertags that turn any line into a structured object with fields; meeting agent transcribes, summarizes, links action items to existing records. [Tana 2025](https://tana.inc/articles/whats-new-in-tana-2025-product-updates), [Medium](https://medium.com/@danielasgharian/tana-review-the-note-taking-app-that-actually-gets-ai-right-639ce08211f3), [aishortcutlab.com](https://aishortcutlab.com/tools/tana/review)
- Fails: near-universally cited huge learning curve.
- Steal: applying a schema to a raw capture automatically gives it fields (status, date, person).

## Memory wearables and passive daily-summary UIs

### Rewind / Limitless
- Pattern: continuous audio capture auto-organized into a day timeline with diarization, searchable transcripts, generated daily summary. [rewind.ai](https://rewind.ai/what-happened-to-rewind/), [Product Hunt](https://www.producthunt.com/products/rewind-2)
- Fails: the standalone product wound down after Meta's acquisition.
- Steal: a passive timeline view of the day as a companion to the structured day page.

### Amazon Bee, Friend
- Bee: hardware praised; real-world reviews found latency, false positives, and a "tiny spy camera" feeling. [bigguyonstuff.com](https://bigguyonstuff.com/ai-wearables-2026-honest-review/)
- Friend: WIRED called it a device that "eavesdrops constantly" and bullies its users with snarky commentary. [techbuzz.ai](https://www.techbuzz.ai/articles/friend-ai-necklace-review-the-129-wearable-that-bullies-you)
- Steal (negative): an AI's proactive personality commentary damages trust if it isn't opt-in and low-frequency.

## Conversational-first shells

### Perplexity
- Pattern: search-first conversational UI with citations; Deep Research produces a structured brief. [G2](https://learn.g2.com/perplexity-ai-review), [datastudios.org](https://www.datastudios.org/post/perplexity-ai-and-voice-conversation-features-live-queries-and-real-time-responses)
- Steal: a structured-brief output mode invoked from chat for on-demand deep dives.

### Sesame (Maya/Miles)
- Pattern: a speech model producing natural pauses, tonal shifts, self-correction. [Sesame](https://www.sesame.com/research/crossing_the_uncanny_valley_of_voice), [PCWorld](https://www.pcworld.com/article/2623695/i-was-so-freaked-out-by-talking-to-this-ai-that-i-had-to-leave/)
- Steal: prosody quality matters for voice check-ins; the uncanny realism itself is the complaint in several reviews.

## Voice-capture tools

### Wispr Flow vs. Otter
- Pattern: system-wide dictation hotkey that turns speech into clean text at the cursor vs. a meeting bot with diarized transcripts. [wisprflow.ai](https://wisprflow.ai/post/wispr-flow-vs-otter-december-2025)
- Steal: hotkey-triggered dictation as the latency-to-capture bar.

## Personal "life OS" projects (GitHub / Claude Code ecosystem)

### seandavi/lifeos-template, davidhariri/life-system, TaylorHuston/local-life-manager
- Pattern: a markdown vault driven by Claude Code skills; a morning routine reads yesterday's journal, today's calendar and email, then proposes 2 to 3 priorities; daily and weekly reviews close the loop. [seandavi/lifeos-template](https://github.com/seandavi/lifeos-template), [davidhariri/life-system](https://github.com/davidhariri/life-system), [TaylorHuston/local-life-manager](https://github.com/TaylorHuston/local-life-manager), [Killer Code roundup](https://cc.deeptoai.com/docs/en/community-tips/claude-code-life-os)
- Works well: independent convergence on "files + scheduled skill + propose, don't decide".
- Fails: text-file-only, no GUI, no mobile.
- Steal: a named, inspectable "propose 3 priorities" step maps to a proposal inbox.

## Design essays: chat vs. structured UI for agents

- Amelia Wattenberger, "Why Chatbots Are Not the Future of Interfaces": a bare text field has unclear affordances and forces alternation between typing and reading. [via UX Collective](https://uxdesign.cc/the-chat-box-isnt-a-ui-paradigm-it-s-what-shipped-96e931d92769). Steal: surface affordances (suggested replies, visible entities), minimize round trips for known-shape tasks.
- Maggie Appleton: most LLM UI should be "spell-check sized" - scoped, embedded features rather than one open-ended chat box. [maggieappleton.com](https://maggieappleton.com/topics/artificial-intelligence). Steal: small single-purpose AI actions inside existing structured surfaces.
- Linus Lee / Geoffrey Litt, "Is chat a good UI for AI?": GUIs win for repeated workflows; the fix is malleable UI combining GUI ergonomics with chat open-endedness. [geoffreylitt.com](https://www.geoffreylitt.com/2025/06/29/chat-ai-dialogue). Steal: chat is the fallback for the novel ask; repeated actions get a dedicated view.
- Allen Pike, "Post-Chat LLM UI": co-authorship (document primary, chat sidebar), generative right-click menus, natural-language search, summarize-and-synthesize instead of chat, "do the obvious thing" next-action suggestion; warns fully generated unpredictable UI is hard to learn. [allenpike.com](https://allenpike.com/2025/post-chat-llm-ui/)
- Smashing Magazine, "Design Patterns For AI Interfaces": task-oriented controls beat prompts for continuous parameters; output visualization via charts and tables instead of prose; embed AI where the work already happens. [smashingmagazine.com](https://www.smashingmagazine.com/2025/07/design-patterns-ai-interfaces/)
- Google Developer forum, "Beyond the Chatbot": five interaction models - prompt-driven, guided conversation, form-based, dynamic action suggestions, ambient/proactive; pick per feature. [discuss.google.dev](https://discuss.google.dev/t/beyond-the-chatbot-designing-effective-agentic-ai-experiences/289245)
- LangChain, "Ambient Agents": human-in-the-loop at three moments - notification, questioning, review. [zbrain.ai](https://zbrain.ai/ambient-agents/)
- Vercel AI SDK, "Generative UI": the model picks which designed component to render from a fixed kit. [Vercel](https://vercel.com/blog/ai-sdk-3-generative-ui)
- Approval-queue pattern: human-readable proposal, risk-tiered actions, queue reviewable at a glance or reviewers rubber-stamp. [eucalipse.com](https://eucalipse.com/articles/ai-agent-approval-queue-human-in-the-loop), [edilec.com](https://edilec.com/blog/ai-11018/approval-screens-high-risk-agent-actions/)

## Synthesis: 12 named UI patterns for a personal daily-work AI

1. **Morning brief as a single card/scroll stack.** Pulse, Gemini Daily Briefing. For: heterogeneous content fits a tappable feed. Against: this exact pattern failed in production (Pulse). Keep the format, generate only from explicit routines, give every card a next action.
2. **Explicit scheduled routines over inferred proactivity.** Gemini Scheduled Actions, Claude Routines, OpenAI's post-Pulse pivot. The pattern that survived. Gap: a visible, editable "your routines" surface.
3. **Proposal inbox / approval queue for agent writes.** Ambient-agent notify/question/review triad; life-OS "propose 3 priorities". Silent writes destroy trust (Motion/Reclaim). Needs risk-tiering or it gets rubber-stamped.
4. **Generative UI: chat renders the real component, not prose.** Vercel, Smashing output visualization, Notion Q&A. Requires a fixed component kit shared by chat and Today.
5. **Chat as command line under a structured page.** Notion Q&A, Tana per-node commands, Cursor-style co-authorship. Scoped chat entry on Day, Journal, Habits pages.
6. **Voice-first capture with async review before it becomes structured data.** Wispr Flow, Reflect, Tana. Minimize steps before speaking; auto-triage into a typed draft to confirm.
7. **Passive timeline of the day.** Limitless/Rewind, Bee. The view is worth having from explicit signal (calendar, done tasks, journal, memos) without hardware.
8. **One-thing hero / single priority surfaced first.** Sunsama daily goal, life-OS priorities. The one convergent pattern across daily-planning products; never auto-decided.
9. **Time-block daily ritual with a visible overload warning.** Sunsama manual-confirm vs Motion/Reclaim silent auto-move. Keep it manual-confirm.
10. **Malleable, task-specific mini-UI instead of one big chat box.** Wattenberger, Appleton, Litt. One-tap structured controls for the highest-frequency actions.
11. **Companion/relationship memory chat.** Dot, Sesame. High engagement, high risk; evidence for staying utility-paired.
12. **Typed knowledge objects as the retrieval and agent-context unit.** Tana, Capacities, Notion. Auto-type captures the way Tana does rather than adding a general notes graph.

## Confidence and gaps

High confidence on the ChatGPT Pulse shutdown causes and on the Motion/Akiflow/Reclaim trust complaints. Lower confidence on aggregator review sites used for Sunsama/Reflect/Capacities/Tana. The UX Collective essay was read via secondary summaries (403 on direct fetch). Wispr Flow/Otter claims come from vendor comparison sites only.