# Clawsy redesign brief (shared by every design agent) You are one of ten designers exploring a redesign of Clawsy. Each designer has a different lens (given in your own prompt). Read this whole file first, then the evidence folder, then design. Evidence folder: `evidence/` next to this file. - `evidence/current-app/*.png` - screenshots of the live app today, desktop (1440) and mobile (390): today, chat, tasks, journal, docs, voice, insights, more, storyboard, plus two full day pages. - `evidence/v3-mocks/*.png` - the July 2026 mocks the current shell was built from. - `evidence/current-data/checkin-2026-09-03.md` - a REAL morning-run day file (JSON block: headline, plan, calendar, reminders, yesterday review, habits, nutrition, news). Use its content in your mocks. No lorem ipsum, ever. - `evidence/current-data/principles.md` - who Clawsy is, voice, what a good and bad run looks like. - `evidence/current-data/clawsy-agent-contract.md` - how the agent actually operates (runs, captures, verbs, write contract). - `evidence/current-data/router.md` - where every captured thought goes. - `evidence/current-data/day-page-redesign-2026-07-31.md` - Gabriel's verbatim direction on the day page. - `evidence/current-data/product-overview-v2-era.md` - older product overview; the core loop and surface list still hold, the DB details do not. - `evidence/current-data/DESIGN.md` - the current color, type and spacing rules. - `evidence/ux-feedback-quotes.md` - verbatim quotes of Gabriel's UX feedback over time (may still be being written when you start; re-check before you finalize). - `evidence/market-research.md` - survey of personal-AI daily-work UIs (same caveat). ## What Clawsy is One user: Gabriel. Solo founder with a day job, ADHD, sporadic energy (multi-week conviction sprints, then low periods). Clawsy is a persistent companion that helps him run his life. Not a task app with a chat box. It lives on an always-on Mac, a web app (local.myclawsy.com) and an iOS shell that embeds the web app with a native 4-tab bar (Today, Chat, Capture, More). The agent (Claude Code) runs three scheduled routines: morning 5 AM (writes the day page: One Thing question, plan, schedule, news, weather, habits, nutrition), afternoon 2 PM (ingests his voice memos and Signal notes-to-self and routes them per router.md), evening 7 PM (triage, one batched question), plus a silent 2 AM janitor, weekly and monthly reviews. Each run ends in at most ONE message that complements the Today screen. He captures by voice (Monologue app on the phone), Signal notes to himself, and chat. Everything is files in a git repo; every write is a commit; undo is git revert. Value test (the agent earns its keep only if it): keeps him on track with his actual goals; explains why things change; protects a low-pressure place to reflect; surfaces honest patterns from evidence; fights diffusion (holds him to one path). Core loop: capture -> extract -> review -> apply -> execute -> reflect -> recover. Entities: day (plan, check-in, recap), task (posture: active / someday / done / dropped, optional surface date, owner can be claude), idea (browsed, not done), lists (reading, watching), journal (dated, agent may only append from a voice memo, never edits), docs/notes (durable), trackers and habits (Google Sheet is the source: chest press, TRX, face pulls, meds, meditation, happiness, vices), nutrition log, goals (one live: "Launch one new product by January 2027"; this week's plan is HIS slot), persons, voice memos (with status unprocessed / processed and a recording kind: monologue / conversation / ambient), notes to self, reminders (attached to tasks), calendar (read via Google), news, weather, portfolio watch. Current surfaces: Today (day page), Chat (per-day thread + saved sessions), Tasks (Triage 47 / Committed 2 / Snoozed 2 / Done 18 tabs, icon row actions), Journal, Docs (file browser), Voice, Change history, Insights, Settings, a floating mic FAB, and a mobile Library (More) page. ## Principles distilled from Gabriel's UX feedback (numbered, cite them in your README) P1 His day, not Clawsy's internals. Never show ids, file paths, column letters, function names, run names, counts of misses, or process talk ("12 agents", "the run did X"). The commit message holds that. P2 Never plan on his behalf. Clawsy does not commit for him. The One Thing and this week's plan are HIS slots; empty says "Awaiting Gabriel", never a seeded suggestion. Runs never create tasks. Anything inferred from a musing is a proposal, held for one batched question. P3 Every item stands alone. If reading it needs yesterday's context, it is written wrong. P4 State the conclusion, hide the data. "Good yard weather until 4" beats "mist, high 63F". Numbers live one tap down. P5 Aesthetics are a requirement, not a polish pass. Images, icons, logos, color. Not bland. Light mode, editorial, one restrained accent. A yard-work day gets a picture and a sentence, not a chip. P6 Minimal AI prose on the page. No headline sentence, no "Yesterday: ..." line, no suggestions, no open-question section. Conclusions, not narration. P7 Never redundant. Do not show the same thing twice ("New task" next to "Create task"; "Completed" over a struck-through task; a message that restates the screen). P8 Explain why things change, with a receipt. Every agent action is visible, attributable and undoable. Suppression is loud, never silent. Failures say what could not be read. P9 One message per run, and it complements Today rather than repeating it. Quiet days are fine; do less, not more. P10 Fight diffusion. One path, one thing. Bounded lists (open tasks capped at 50). Ideas are browsed, not surfaced daily. P11 Habits, goals and plan are three different things. Habits are inputs, not prose. Goals are standing commitments with progress and the current plan. Plan is only what he committed to today. P12 Capture is frictionless and routed exactly once. Nothing evaporates. Provenance survives (you can always get back to the memo that produced a task). P13 Low pressure. No grades, no scores, no lectures, no streak shaming. Recovery (stale, dropped, snoozed) is a normal workflow, not a shame event. P14 Mobile is first-class, not a stacked desktop. 44pt targets, buttons honor their text size, no horizontal page scroll, the mic is always reachable. P15 Shaped for ADHD. Lead with the next action. Few items per decision. Progressive disclosure. Visible progress. Working memory is small: the screen restates state, it never assumes he remembers. P16 Don't wreck what works. Section order and slots on the day page were approved 2026-08-10; the mic FAB, the light theme and the 4-tab mobile shell are settled. Propose changes to them only with a reason tied to a principle above. P17 Taste anchors: Things 3 calm plus Apple Fitness boldness. A calm typographic base with ONE bold hero moment per screen, not uniform loudness. His words: "No soul. Feels like a dashboard." and "janky / AI generated" are the failure modes; "clean, bright and professional" is the target. P18 Tap to expand, in place. A task, memo or event opens as a card or modal where he is; it never navigates him away to a file or a different surface. Links to entities show an icon plus the title, never an id or a path. "When I say a card, it should be a modal." P19 Today chat is the primary chat entry. On desktop it opens beside the day (content shifts, chat takes the right half, only when opened); on mobile it is one tap away; he narrates over the app by voice and the day updates. The general chat must not get dirty with check-in traffic. P20 Captures are reviewed conservatively and explicitly. He can see EVERY voice memo and note-to-self not yet accepted, in one place, with date and a generated title, and he checks or dismisses each. Processing skews conservative: five bullets he said, never a ten-page document he did not. P21 Data is live from the canonical store. Habits, nutrition and metrics read the sheet and logs at load time; nothing is baked in. Logging is a plain dialog with one Log action and shows the running total so far today. P22 A control looks like what it does. No bare glyph that silently snoozes a task. Hover and pressed states on everything. A heading that shows a count is the toggle for that count, and the count means what it says. ## What to produce (all files under `directions//`) 1. `README.md` (bullets, under 150 lines): the concept in three sentences; who it is for (him, at 5 AM, at 2 PM on the phone, at 9 PM tired); information architecture (surfaces, navigation, what moved and why); the component inventory (every new or changed UI component, one line each); how each principle P1-P16 is honored or knowingly bent; what this direction would break in the current app; what it steals from which product (URLs); open questions. 2. `storyboard.html`: a web storyboard of one real day, 5 to 7 frames laid out left to right or top to bottom with a caption under each: 5 AM the morning message and Today; 9 AM he sets the One Thing; 11 AM he records a voice memo while walking; 2 PM the afternoon run routes it and something appears; 7 PM the evening question and triage; 9 PM he reflects; plus a Saturday weekly frame if your lens needs it. Each frame is a real drawn screen (HTML/CSS, phone-sized where it is a phone moment), not a description. 3. `mocks.html`: full-fidelity screen mocks. Required: Today on mobile (390 wide) and desktop (1440), Chat, Tasks or whatever your lens replaces Tasks with, and ONE surface that is new in your direction. Use the real content from checkin-2026-09-03.md and the task titles visible in the tasks screenshot. 4. `shots/*.png`: screenshots of your own HTML taken with `tools/shot.mjs` (see `tools/README.md`). You MUST look at every screenshot with the Read tool and fix what is wrong before finishing. Check both 1440 and 390 widths. Rules for the HTML: - Self-contained: inline CSS, no JS frameworks, no build step. Google Fonts allowed. Images: use CSS gradients, inline SVG, or emoji-free icon SVGs; you may reference the existing watercolor hero via a soft gradient placeholder labeled as a painting. - Light mode, editorial, one accent. Body 15 to 17px. One spacing scale. Real 44pt tap targets on mobile frames. - No em-dashes or en-dashes anywhere (use a plain hyphen or restructure). No lorem ipsum. No "v2", "draft", "mock" labels on the page. No process narration ("this mock shows..."): captions describe what HE sees and does. - Frames render at true size inside the page (a phone frame is 390 CSS px wide); the page itself may scroll. ## References you should actually look at - Try the Mobbin MCP tools (search_screens / search_flows) once. If they error with a paid-plan message, use WebSearch and WebFetch instead: mobbin.com public pages, real app teardowns, Dribbble and Behance shots, product sites. Cite every reference with a URL in README. - Products worth studying for your lens: Poke (interaction.co), Dot (new.computer), ChatGPT Pulse, Sunsama, Amie, Reflect, Apple Journal, Things 3, Linear, Raycast, Arc, Notion Calendar, Rise, Structured, Gentler Streak, Bearable, Cron/Notion Calendar, Superhuman. ## Bar A stranger who has never met Gabriel should be able to read your storyboard and understand the product. Gabriel should look at your Today mock and see HIS day, with nothing he did not commit to and nothing that needs the system's internals to parse. If your direction is just the current app with rounder corners, start over.