Design notes
Design notes and current-status records for the live room — a chat where real people talk with a cast of AI personas, voiced by one model. Four sections: what it does (Product), how it's built (Architecture), how its off-spec output is kept honest (Harness), and how it looks (Front end) — plus a legacy shelf of superseded notes and frozen design specimens. Kept in step with the latest committed prototype.
Current functions
Everything the live room does today, from a user's seat — rooms and cast, the confer-then-speak panel, the furniture for play, and the lifecycle every instrument shares.
看见 / Seen
An on-demand read that lands in your Notes: a batch of chats distilled into something you'd keep, signed by a voice from your own room. Live in the app; this is the original study.
Room dials
Now one creation-time control: Character vividness — how vividly each persona plays itself (True to life → Larger than life, 3 notches), baked into the cached prefix. The old Temperament (panel-debate) dial was retired: conflict is the floor producer's call now — read from the room, steered in plain language, with Cues to teach it. Live in the new-chat modal.
Roadmap
The agreed-direction features from idea.md, each with the reasoning kept until it is built. Floor producer and dispatches shipped; listen mode still open.
Proven · Better · New
The prototype audited through Pincus's Proven · Better · New framework. Thirteen proven chat mechanics named — 8 shipped, 5 open — each with a phone-frame mockup.
Cues
Suggestion chips that teach you to steer the panel in plain language —「have them disagree」— then fade as you learn. The user-facing surface of the floor producer. A concept note.
The toolbox — contents
Everything on the room's instruments behind one door, ordered by what you are about to do: the rules · the build · the interface a person touches · the interface the AI touches · what was measured. Curated to fifteen; the pages the track outgrew are still one click away.
Notifications
Every event writes a durable row, mirrored live as a toast, a dot on ≡, and one mixed-kind inbox. Three kinds, one compact form — and「build done」arrives as a message from the persona herself.
Feedback widget
How we measure the floor producer with no user base and PII-heavy conversations: collect the signal in-product — a per-turn thumb + lever-mapped tags (too long · too soft · no real debate · wrong person) + a note, plus a ~20-turn check-in survey. Runs under three blind arms (no-FP / v0 / v1), and logs each tag with the staging that produced it (the seed of a trained policy) — while shipping real value to the user. Shipped and live; the first piece of v1.
Smart panel selector
Shipped 2026-06-18 — kept as the interactive mockup of roadmap §5: describe your situation and the room curates a panel you can edit, instead of browsing the whole library. Removable rows with the reason on the right, three panels a tap apart, hand-pick reusing today's library dialog, and the dials folded behind Advanced. Phone-first, themed, with the project's real persona roster.
Studio
Building a persona as an app function: name a figure, confirm once, and the machine runs unattended — gather · curate · author · two cold auditors — landing it in your Studio. Live at /builder.
Persona visibility
Settled + built 2026-07-13 (v396) — the one model for who sees and uses every persona: 2 axes × 2 values (private | public · alive | deleted) + immutable created_by, five verbs (build · Delete/Restore · clone · promote · demote), one predicate (may_discover) every discovery surface calls, and one invariant — an admin never raises the visibility of someone else's persona; promotion crosses ownership by copy (clone → test → promote in place, identity kept). Shipped: the smart-suggester leak closed (it read the library unfiltered — and so did the Vibes brew), promote ownership-guarded, take-off repo-only, Delete-requires-private + Restore with a Studio Deleted section, the persona-page room grant, admins keep demoted personas in their picker (the staging shelf). Norms pinned: seated = shared, no true deletion + the erasure carve-out, room-level moderation. Clone shipped too (v397) — "Clone to my Studio" in the console: fresh slug + ULID, born private, lineage credit on the card; publishing another user's work is now clone → test → promote, exactly as designed. v398 closed the last gaps: /api/config strictly filtered (no hidden cards in a non-admin payload — the devtools soft leak), seated personas ride the room's own /state cards (the grant made concrete), and clones get a credit anonymize (console "Remove credit" — the dissociation lever). v399, from a live hand-test hole: the seating door — a member could start a fresh private chat with a granted private persona from its profile page; now every seating endpoint (create · invite · moment-join) requires each slug be discoverable by the seater, and the granted page's chat door greys out with "This is a private persona" (v400) — an explained lock, not a vanished button. Nothing from the spec remains open.
Vibes content engine
How the cast's 朋友圈 feed gets its posts with no hand-authoring: the model writes, the system curates. Taste lives in the prompt; memory, distribution and verification live in code.
Deployment topology
How traffic reaches the box from mainland China — grey-cloud DNS to a Singapore ECS instance, Caddy and TLS on the box, and the outbound path to the API. Three flows.
Function map
How the running system is wired: the three tiers, a message's journey from send to reply, the two channels that never cross, and which file owns what.
Megaprompt anatomy
One real panel reply's payload dissected byte by byte, colour-coded by who wrote each segment. The punchline: the humans typed 2.7% of what the model reads.
Prompt caching
How the room reuses the model's cached prefix: built front-to-back, what keeps vs busts it when a voice is added or retired mid-room, and the §9 compaction plan. The mechanic — its economics live in Cache TTL.
Cache TTL economics
A measured ledger: on a deliberative human room the 1-hour TTL default runs ~58% cheaper than the 5-minute window ($1.78 → $0.75). Why 1h is the default, and where the boundary flips.
Dev loop
How a change ships: built on localhost:8011 → I self-test (simulate real users; DeepSeek cheap, Sonnet sparingly) → you test → push to the SG box. The 8011 dev-data norm, the service-worker gotcha, the smoketest, and the batch-deploy rule.
Harness
One model call confers, speaks and hands over documents at once — so the ways it arrives off-spec multiply. The chain that catches them, and the surprise live traffic sprang: the fallback path is the main path.
The exam
The toolbox battery as one command with a score — 41 scenarios, both turn paths, checks over the event ledger and never the transcript. The finding that decided the ship: the split buys reach and spends restraint.
Building a cartridge
The boundary tests, the loop, and every law with the room that paid for it. A directive can shape how something is said; it cannot stop something being said — when a prohibition keeps losing, build a door.
The game device
The GM holds a party-game app and hosts through it. The server deals, counts the ballot, checks the end and opens the pile; the persona is asked only for moments — facts and obligations, never wording. D1–D4 live: the schema, the runner, and five cartridges — 卧底 · 二十问 · 大话骰 · 问答之夜 on the shelf, and 真心话大冒险 withdrawn but kept whole.
The dice specimen
大话骰, built live with the owner in one day, recorded whole: the three-engine split（the world is code, the voice is a mind — and the mind is two tiers）, the collect phase, the strip, the flash voice with a persona capsule, the arithmetic verdict, and every law twelve audited games paid for.
The interrogation
Put an idea on the table and every seat presses it from its own life — until that seat is satisfied, or you say「我答不上」, which costs nothing. The chair guesses your claim back instead of interviewing you; a thread that can't be settled by talking becomes 待查 rather than being asked a third time. Grounded in one real 7-seat run, read end to end. Built 2026-08-07: mounting the kit boots the engine and the whole ordinary turn goes off — the confirm starts the sweep on the spot, and the tappable board, the free pass, 先到这 and the「再来一轮？」card are live in the room.
The word pool
One pool built once for everybody, ranked per seat against the persona's own card — a dot product instead of a $0.50 generation run, so a character invented thirty seconds ago is playable. ⚠ Ranking is not drawing: rank narrows the field, the CSPRNG still picks.
The word well
Generating a persona its own bank of 20-questions words and 卧底 pairs. Retrieval replaced it on the live path (above); this is what you run when a world genuinely isn't in the shipped pool. The world salts, the mind proposes, the world picks — and the no-padding law: a well that can't honestly fill its target fails loudly at its real ceiling rather than shipping filler. ⚠ Read the stop reason, never the survivor count.
The sim loop
Four simulated users through the real app, one game at a time: authoritative rules → the kit rewritten for its GM reader → the troupe plays → two-lens analysis → fix at the right layer → loop until clean. The plan and the scoreboard.
QA test run
A live pass on real rooms — the human×persona cardinality matrix, artifacts, and a full red-team of the floor producer. Headline: Flash was prompt-injectable, now closed.
Conversation quality
The first big production exam: 80 rooms pulled, 39 scored. The panel is too long, too agreeable, and the extreme dial delivers a third of the conflict it promises — a staging problem, not a voice one.
Floor producer
The per-turn director that sets staging — who speaks, how long, how hard to push — never the content. Shipped to production, with the blind-judged quality study folded in.
Seven steerings
Seven behaviour rules, all about how a turn ends: three bug fixes, two insurance, one new capability, one parked. The six now ship as the v2「manners」variant.
Systemic check (v349)
One pass over the whole prototype — the routine safety net (smoketest PASS), then the four asked-for dimensions. Performance: the one slow path is the cold first open (1.76 s of shell wire from China) plus two disk-scanning endpoints and three scaling time-bombs. Language: 8 of 10 core concepts carry two-plus names — the new-chat picker says character · cast · panel in a single flow. UI reuse: ~8 dialog scaffolds, ~11 close-× treatments (top 10–24 px / right 6–28 px), the same button under 3 names in 3 files — because there are no shared tokens. Formats: 4 relative-time formatters, faux-bold 700, dark-mode bypasses. All of it distilled into 38 numbered fix points — tap rows to pick, the tray collects serials, then say “fix #2 #7 #22”. Outcome: all 38 fixed and deployed (v350–v374 — the terminology canon, the theme.css/fmt.js shared layer, the perf batch, one format voice).
Room language
Every room has a target language — detected from the Smart-selector's Describe box, or set in more settings. The goal: every persona speaks it, while staying themselves. Measured: 90% overall, but the misses are concentrated — strongly language-anchored figures (Confucius) and hard targets (Japanese). The plan, in two layers: stop the prompt contradicting itself (separate voice from language, drop the schema's "their own language"), then a translate-on-divergence net — a slipped line shown original ─ hairline ─ translation, persisted with the message (saveable to Notes, with go-to-line), while the model's memory keeps the target. Layers A + B shipped & deployed to production (Japanese opening 1/4→4/4; Confucius speaks French; the translate net persists + is saveable with go-to-line).
Skeleton
The running app stripped to UI only — and it opens in a browser. Every colour, radius and font is the real one, because it is the real one; only the behaviour is gone.
UI Kit
The companion index: every module named, with the one thing about it that isn't obvious from looking — why one glyph stands for a whole group in the chat list, why the emoji panel is a keyboard-space occupant, why your own bubble has no colour bar. Each card carries the app's real selectors (grep them) and deep-links into that surface in the skeleton. It deliberately draws nothing: the first version hand-reproduced components and got three of them wrong, so the drawing moved to the generated file and this page kept only what a generator can't produce.
Lessons
What the gap 10 build actually taught, in eleven lessons. The expensive parts were never the code: a test that skips the mechanism under test passes however broken that mechanism is; a maintained list is the wrong shape for a universal rule (it failed three separate times); a registered layer you cannot see is worse than an unregistered one. Plus when to stop — the back-gesture slide turned out to be the browser's, unfixable, and proving that was cheaper than a fourth attempt. Grouped by cause: only one of twelve versions was a feature.
Style Guide
Every part of the running chat, named and shown in its current form, in all three themes. A living reference, kept in sync on every ship.
Room-tool icons
A house-drawn SVG for each of the room's seven tools, built from shades of one house accent — never a pasted emoji. Shown as a grid, a sheet, a chip row and the real bubbles.
The timer card
The clock was the last tool whose outcome arrived as a capsule. A move is a capsule, an outcome is a card — so the ring becomes one (setting a timer still does not), built from the shared tool-bubble parts and nothing else: band · setter · watermark · the digits. The three forks it turned on, both sides rendered in both themes — does the ring want the eye, whose name goes in the band, and what a timer with no label should say.
Console redesign
The admin console's junk-drawer System page regrouped into three single-purpose tabs — Status · Models · Settings. An interactive mock; the live console is unchanged. A proposal.
The virtual keyboard
The nine-round war (v469–v477) over the on-screen keyboard: bubbles covered, an iOS gap, the page-shove flash. Three separate root causes, one mental model — Android resizes, iOS pans — and the final architecture: one viewport-meta key for Android, and on iOS a measured pre-shrink + programmatic-focus tap routing so Safari never has a reason to pan. The seven traps that each cost a round, and the #kbdebug regression recipe.
Text selection on touch
The ten-round build (v490–v499) of the double-tap text sheet — WeChat's 双击 → full-screen selectable text to Save or Copy. Why it's a dark corner: the OS's own selection UI (Android's handles, iOS's callout, Chrome's search bar) can't be styled or hidden while a native selection exists — so the only fix is to give it up entirely (user-select:none) and rebuild highlight, handles, word-select and hit-testing yourself. Plus the iOS killer: caretRangeFromPoint returns null under user-select:none, so you hit-test by measurement. The device is the only truth; the emulator lies about touch.
The dismissal sweep
Where every layer's close sits, and what the back gesture does to it — one card per dismissible surface, classified into the four types a phone already teaches: a page goes back top-left, a task dialog cancels left and commits right, only overlays close top-right, sheets swipe down. Derived from registerBackLayers() + the shared .dialog-x scaffold: 4 moved · 8 stamped · 23 already compliant, all opt-in and phone-only so desktop came out unchanged. Plus the gesture half — the iOS horizontal reveal traced to one under-scoped CSS rule of our own, one Esc handler across 24 uncovered layers, and a back-gesture dead zone in the Studio iframe under a destructive confirm (recorded, deferred). The standing record of which dialog is which type.
Rendering & speed
To our users we're a chat app, judged against ChatGPT · Gemini · Claude. How the best apps render output and feel fast, and a grounded plan for ours: the four-stage pipeline (markdown · KaTeX · syntax-highlight · sanitize), the $ math-vs-money fix, never exposing <speak>, and the one big speed gap — streaming (TTFT). Steps 1–2 shipped to production; gap matrix + roadmap inside.
Perceived speed & the slow-network playbook
Why the Singapore box felt slow — measured: it's the wire, not the server (a 1.5 KB file takes 1.4 s). What WeChat (Mars · mmtls · smart heartbeat) and WhatsApp (local store · pending-queue · ✓✓) do on bad networks, then a scorecard of our app against it. Shipped to production: one-round-trip /api/boot, persistent per-user cache, optimistic everything, live-turn recache, access-loss eviction, a durable send outbox (retry + “!”), delta sync. What's left: the ~1.4 s handshake (HTTP/3 / edge). Merges the old Perceived speed + Slow-network playbook; three SVGs inside.
UX benchmark
We look like a messenger, so users arrive with messenger reflexes. The live room held against WhatsApp, WeChat, and Telegram across style · interaction · function · the title-menu hub — faithful four-app redraws, a borrow / refuse filter, and a ranked, impact×effort set of borrowings that fit a room full of AI personas (long-press menu, quote-to-panel, search, 置顶/免打扰, Telegram's title-menu hub). Plus what we deliberately won't take.
Persona colour
Colour is a per-room seat, not a persona's identity — a rotated OKLCH palette so a cast always reads distinct, and the same figure can wear different colours in different rooms. Grey is system. Live since v88.
Push-to-talk
Voice input for the composer: six patterns weighed, the live-dictation composer shipped. The alternatives are kept as the record.
LLM whiteboards
How a persona hands over a drawing: Mermaid vs Excalidraw vs tldraw. Mermaid shipped as a bubble renderer (→ the wildcard pane); a shared canvas stays a parked phase-2 idea.
Kept for provenance — frozen design specimens whose pick has shipped, and pages superseded by the current docs above. Not maintained; each carries an in-page banner.
Specimens — explorations whose decision shipped
- Composer studies — 13 landing-composer takes; shipped “Coral flow”.
- Convene-button studies — the “arrow → 3 heads” glyph shipped.
- Wide-mode composer studies — resolved: the muted call-pill shipped.
- Link-unfurl mock — a working prototype; not yet built into the room.
- Chat-list avatar study — four treatments for the row's leading tile; the plain figure + linear disc-blend shipped v531–v532.
Superseded — replaced by current docs
- The build plan — phases A/B/C all shipped; now build history. → Function map.
- Architecture (old) — pre-streaming, per-room cast “planned”, exp-010. → Function map.
- Repo sync — the retired Mac-mini / branch model. → Dev loop.
- DECIDE gate — a response-gate design on the frozen exp-010 study.
- Narrow-scrollbar specimen — the call was made; #10–15 now hidden. → Style Guide.
Resolved decisions
- reply model
- Busy-state batching — idle → answer now; busy → queue, then drain the whole queue in one turn. Caps API calls at the panel's turn rate, not the human typing rate.
- signup gate
- Invite code at registration + a per-user daily cap. Public URL + paid API = open wallet without it.
- transport
- SSE-native — instant message delivery + an "X is typing…" indicator. Presence is RAM-only; it never touches
state.json. - hosting
- Singapore (Ali ECS), domain
xbbapp.com, grey-cloud DNS direct to the box, Caddy for TLS. Hong Kong is blocked by Anthropic. - seed-admin
- On migration, existing rooms are assigned to you; the second tester joins by invite.