Ink 墨 — the persona newsroomconcept

Personas write long-form essays on what's happening in the world — carefully, in their own tone and thought process — and the essay feed becomes the app's entrance. Product name: Ink (墨) — chosen 2026-09-01, replacing the placeholder「Vibes」; internal slugs stay vibes-*, and the app's shipped short-post surface keeps its old name until its own rename ships. Companion pages: Vibes (the extractor / writer machinery this builds on) · Essay specimens (the first live run — three real essays, 2026-08-31). Settled in discussion 2026-08-31. Shipped so far: v1038 the Ink root tab (first of four); v1039 the static edition — /ink/edition.json + hero art served as content-as-files, rendered in the researched sequence (dateline · lead ink-band · Fresh Ink rows · the Ink Well · end marker), Playfair display + Source Serif text, bylines as doors to the persona profile. Step 1.5 next (owner's contract): the reader's door seeds a room with the essay as opening context, and the in-app article link is a SPECIAL BUBBLE — the unfurl-card grammar (rubric · title · dek · hero thumb) as its own bubble kind; tapping it opens the Ink reader. The door drops that bubble as the room's first artifact, so the essay is quotable furniture in the conversation it started. The brew, the gate and the triggers remain unbuilt.

The problem

-ish is a lean-forward product (companion-market, difference #8): users deliberate before speaking, and deliberation has a start-up cost. The blank composer is where that cost is highest — nothing to react to, no thread to pull.

The obvious fix — make personas proactively raise topics in chat — has a known failure mode: an eager persona guessing at interests reads as noise, and being talked at is worse than silence. The persona shouldn't push a topic into your room; it should publish its thinking where you can browse it, and let you carry the piece that catches you into a conversation.

The shape — settled

• The cast is the newsroom. A day's worth of fresh pieces every day — because interest-matching across a diverse user space is unpredictable, the list must be long enough that everyone finds something. But no persona is on a treadmill: thirty seats × ~2 pieces/week ≈ a dozen a day. Volume across the roster, scarcity within each voice.
• Reflection, not commentary. The world event is the doorway, not the subject: the piece walks from the news into the persona's own territory — the China-policy seat on what the newest property measures mean on the ground; Feynman or Einstein on what this month's physics result actually changes. Reads WeChat-article long: several careful paragraphs, actually worth reading — never a one-line hot take. (For real-figure seats this is also the safety posture: their method applied to the event, not invented opinions put in their mouth.)
• Everyone's essays in the window. The feed shows all personas' pieces for the period — browsing is how users meet personas they haven't friended. Personalized reordering (from chat topics + click behaviour, TikTok-style) comes later, when volume outgrows a screenful; -ish will have a ranking signal feeds envy — users converse with the authors.
• An automated floor, not editorial scarcity.「Written by an AI」is no longer the bar — half the WeChat feed already is. The differentiator is a persona with a position, a biography and a thought process, and that is precisely the hard part at volume — so it is measured, not assumed: the AI-tone meter / de_tell net become a publish gate. A piece that fails reads as generic AI → the persona simply didn't write today. Run-2 finding (owner + CC): the gate must meter both languages' antithesis reflex — zh 不是…而是… AND its English twin「not X — Y」, which appeared in all five English pieces — scoring density per piece and sameness across the slate (every author sharing one rhetorical skeleton is the model's fingerprint, not the persona's).
• Every essay ends in a door.「Sit down with the author」— one tap opens a room seeded with the essay as opening context. The feed is not a content business; it is the on-ramp to conversations the persona has already prepared for.
• The economics work. An essay is written once and read by every user — unlike chat, cost does not scale with readers. Drafted on the cheap tier in the daily brew (off-peak), gated, published.

The feed — first cut

The 小红书 / Pinterest grid: 2-col masonry cards, cover + title + byline, detail on tap. The layout is fully familiar to Western users (Pinterest taught it for a decade; the Jan 2025 RedNote moment and Lemon8 reinforced it) — but in the West it codes as visual lifestyle discovery, so essays must win the cover slot: the persona is the cover. Portrait/monogram + bold serif title — each card sells the essay and introduces its author. No per-article generated art.

Left: the entrance —「Today」is everyone's pieces for the period (the browse-to-meet-personas surface);「Following」filters to your personas;「Rooms live」shows rooms that grew out of essays. Right: the article — persona-cover, byline card with Follow, the「wrote this after」provenance line, several real paragraphs, and the door at the end. Card covers vary in height (masonry), tags are the persona's territory, read counts build the newsroom feel.

The formats — one register is a magazine, eight are a feed

The three specimens are all one genre: the first-person reflective essay (the signed column; the 公众号随笔 register — news peg → personal experience → argument → quiet ending). A dozen of those a day reads like one magazine. Variety comes from formats — and format is a likeliness, not an assignment (owner, 2026-09-01): every persona can attempt every format, background sets the odds. Munger writes letters often and fables rarely; Plato writing an advice column is not a mismatch to prevent — Socratic advice would be genuinely novel, and this is where -ish beats traditional publishing: an augmented persona can leave its franchise cheaply, a human columnist can't. Don't seek the mismatch, but when it happens naturally it's a blessing. The one invariant: the format bends to the persona, never the persona to the format — Plato's advice column comes out dialogue-shaped, and the moment it reads like a generic agony aunt the tone gate should fail it. Bonus: format variety gives the masonry grid its visual variety free — different card types, no hero pictures needed.

The first-hand boundary

The boundary doesn't kill the experience-shaped formats — it forces each into its honest variant:

• Memory, explicitly dated. Bourdain can't taste the new place; he can write about the market he actually knew, pegged to today's news that it's closing. The dispatch example above is written this way on purpose. And for the history greats this variant is a superpower, not a workaround (owner's example): Da Vinci couldn't attend the Venice carnival of 2026 — but he can read the news about it and compare it to the carnival he actually walked through five hundred years ago. A living columnist's memory reaches back thirty years; the roster's reaches back millennia. No human publication can print that comparison.
• Analysis of the record. MKBHD can't review the device; he can review the launch — keynote ratios, spec sheets, other people's attributed reports — or write the「what I'd test」checklist of skepticism, with no claimed hands-on.
• The confession of distance.「I haven't held it, and here's what the spec sheet lets me say anyway」— stated distance is itself voice, and readers trust a writer who marks the edge of their knowledge.
• The room inversion — the n:m answer. The users did taste the restaurant and hold the gadget. On first-hand pegs the door flips: room → essay instead of essay → room. The persona hosts the people who have hands and taste buds, then writes from the room (the demand peg, trigger #5). The humans bring the senses; the persona brings the frame. No other publication can do this either.
• The boundary governs claims, not confessions (owner, 2026-09-01, after run 2 opened all six pieces with disclaimers). The public record is presumed accessible — a published speech, a broadcast match, a printed report need no「I wasn't there」: Munger's essay IS the response to the speech. Disclaim only where the piece would otherwise imply physical presence (a dispatch, a hands-on review) — and even then as information, not legal notice:「I did not pay attention to this island before」beats「I have not seen this island」. Prompt-craft note: run 2 over-deployed because the rule sat in the editor prompt, the writer prompt AND the angle — an instruction present everywhere gets deployed everywhere (the persona-pipeline lesson, replayed).
• Vantage: alive now, remembering then (owner, 2026-09-01). History seats write as themselves today, reaching back — never datelined from their own century. It is the only vantage consistent with the app: the same Da Vinci must walk into the room when the door is tapped, and he cannot do that from 1518.

The anecdote budget — repetition at volume

The concern (owner, after run 2): profiles are compact — a handful of corpus anecdotes each — and every run-2 piece spent marquee ones (the Dardanelles, Baghdatis, the Central Park stall, the wax heart). At ~2 pieces/week, a persona's ten best stories are ash within months, and later essays repeat material and tone. Four mechanisms, three already proven in-house:

• The used-anchor ledger. The Vibes engine already keeps one for posts; extend it to essays — every deployed anecdote/quote logged per persona, and the writer prompt carries「already used — do not reuse」.
• The anti-echo block. Feed the writer the persona's own last N essays:「don't repeat their anecdotes, structures, joke-shapes or endings」— the exact mechanism that killed repeated closers in Vibes posts.
• Method over memoir. Ration biography: the marquee anecdote is a spice, not the fuel. Munger's run-2 piece runs almost entirely on method applied to fresh material — and method is renewable where biography is finite.「No anecdote in this piece」should be a normal commission.
• The corpus grows in-world. The growth track and the essays themselves generate new referenceable experience —「as I wrote in September, and was wrong about」— so a persona's own published record becomes fresh material. The 商榷 format above accelerates this: the feed starts feeding itself.
• The memory bank — expand the research, outside profile.md (owner, 2026-09-01). profile.md is compact because it rides whole into every chat context — but nothing stops a sibling store of researched stories that never loads whole anywhere. The machinery half-exists: vibes_extract already distills profile + corpus into a voice card + anchor bank (mine mode for famous figures, strict verbatim-cite mode for supplied/built). Grow that bank with periodic research jobs, then deal, don't dump: the essay writer gets profile.md whole (the voice) plus a dealt hand of 2–4 anchors ranked against the peg (the wordpick rank-then-pick pattern) and filtered by the used-anchor ledger. The bank can hold two hundred stories; the writer only ever sees the few that fit today and aren't burned. Provenance discipline is non-negotiable — the bank must never accumulate invented biography, or the first-hand boundary leaks in through the side door. ⚠ Open coherence question: essays are public in-world, but a room not seeded by the essay knows nothing of the bank — a user quoting a persona's own essay at them mustn't meet a blank stare;「the room must not contradict the feed」belongs to the memory track.

Where topics come from — the trigger taxonomy

The budget meeting. A front page is a portfolio, not a pile: the editor sees every pitch at once and composes the mix. So production runs one editor call per day — all pegs from all eight triggers + the roster in, the day's commissioned slate out — which is also where dedupe, beat balance and the ≤2-per-persona-per-week cadence get enforced. (This revises the specimen harness, which ran three independent editor calls.)

The daily mix to aim at, ~a dozen pieces: ≈4 breaking · 3 calendar · 2 demand/reply · 2 appointment columns · 1 wildcard — roughly a good paper's front-page ratio; only the breaking slice depends on the news gods cooperating.

The loop

Open questions