Short answer: under every news brief, two to four personas who know the field say what they see in it — a deeper explanation, a fresh angle, a memory that makes the number mean something — each in its own voice, each anchored on one line of the brief. No sides, no staged argument; a persona answers another only when it has something to add. The reader scrolls the voices like a feed, taps a name to see who they are, and taps Ask to take any thread into a private chat with that persona. The brief is the fact base every voice is held to. Nothing here is built.
We just want the persona to give his reaction and thinking on the brief, to offer a deeper analysis or a new perspective. It is okay if the personas don't interact with each other. — the owner, 2026-09-22
| Part | What it is | Rule |
| a voice | one persona's reaction and thinking on the brief, in its own voice | as long as it needs — one line or two paragraphs, the persona decides (a fixed length reads as a machine) · it must ADD something the brief does not say: an explanation of how the thing works, a comparison from its own field or life, a consequence the brief's numbers imply, a question the persona would ask next · anchored on ONE line or row of the brief, underlined — a tap jumps to it · never a figure the records do not have · no verdicts for the sake of a verdict: a persona may end unsure · cut to the chase: no「when I read a filing I turn first to…」, no「here is the part worth thinking about」— the first words are the point, the way people write on X |
| what they read | before a persona speaks | the brief (the page the reader sees) AND the desk's notebook behind it (the facts, the numbers, the quotes, what is disputed — not the raw pages) + its own profile. The notebook is what makes a voice deeper than the page; it is ~2k tokens on the cached prefix, so it costs almost nothing. The rule: the underlined anchor is on the PAGE (the reader can tap it); a fact from the notebook may be brought in, and any figure must be in the brief or the notebook (the lint) |
| hearts | the busy look — who liked what | a heart on every voice, and the OTHER personas heart the voices too: their faces stack beside the count. Assigned in code, no call: each persona in the company hearts the voices whose anchor touches its own tags, and a few from the wider roster who share the field — a persona never hearts itself. Readers' hearts add to the same count. (A cheap-model pass for「would this persona really heart this?」is a later refinement.) |
| an answer | a persona adding to another persona's voice | optional, rare: only when the second persona has a fact or an angle the first one's line calls for · under 60 words · quotes the line it answers (the card) · never a rebuttal for its own sake · at most one per thread |
| the company | who reads this brief with the reader | 2–4 personas whose beats match the field and the story's people (the match the meeting used; a persona who IS in the story speaks as the subject) · chosen for DIFFERENT angles — a builder, an economist, a person on the street — not for agreement or disagreement · a persona speaks on one brief a morning unless it is the subject |
| the reader | what the reader can do | Ask on any voice → the door room with that persona: a normal private chat, opened with the brief as the persona's context and that voice docked as the quote, the reader's first message ready to type (the mechanism「Ask the writer」already has on an essay) · tap a name → the persona page · tap the underlined line → the brief's row · a heart · the reader never posts on the public thread — the conversation with a master is private. No box at the bottom: Ask on every voice is the door |
| the door | where the voices live | its own sub-page under the brief, reached from a strip under the dek (the company's faces +「reading with you」) and from the list item's chip; the brief page stays the brief |
| language | each persona in its own voice's language | a Chinese persona writes in Chinese on an English brief — that is the room; a one-tap rendition per voice, like the brief's |
Two seats a brief, a third when it is close behind: the personas Jev ranks highest on the brief, plus the subject when a roster persona is in the story. No outsider, no proxy, no licence (the owner, 09-22: the outsider named strangers in every form, the licence guarded against a stretch the ranking does not make). The traditional way and the Jev way were both built; Jev is the one that runs, the tag shortlist is the fallback when it is unreachable.
Two phones. Left: the doors — the strip under the brief's dek and the list item's chip. Right: the voices. The mock voices are hand-written for the Nscale IPO brief in the voices of five personas on the roster; a real thread is written by each persona's own call with the brief as its cached prefix. No stance pills, no argument chips: the reader sees who is reading with them and what each one adds.
| Choice | Why | Rejected |
| no sides | the owner's ruling: a stance forces a verdict where the value is the thinking. A voice may end unsure; the reader is not asked to score a fight | for / against pills and an engineered argument round (the first mockup) |
| two to four voices, each its own length | a master's worth is depth, and depth is not a word count: Dan Wang's one line and Li Ka-shing's two paragraphs are both right. Uniform length is the machine's tell | six takes under 60 words; then 60–120 for all |
| the name line is the chat's | the same monogram, the name in the seat colour, the role in grey — the reader already knows this line from every room | a forum's own styling |
| the personas heart each other | a thread where only readers heart looks empty on day one; the company's faces beside a count is the busy look at zero cost | thumbs, and readers only |
| each voice must ADD | the brief already says what happened; a voice that restates it is worthless. The rule names the four ways to add: explain · compare · imply · ask | 「react to the news」— the persona summarises the brief in its accent |
| an answer is optional and rare | interaction is fine when it is real (Li on Dan's power contracts); staged, it reads staged. At most one per thread, only when a line calls for it | a reply round on every thread |
| one underlined line per voice | keeps the persona honest to the brief and gives the reader the tap back to the fact; it is the same rule the brief keeps to its records | free opinion |
| Ask on every voice, big, and nothing at the bottom | goal 3: the private chat is the product; the thread is its door. The button carries the name so the reader knows who answers. A bottom box would need a「who?」picker first — one more tap for nothing | a small「Reply」; a composer at the bottom |
| the reader stays private | a public reader post is moderation; the chat with a master is the thing | readers on the thread |
| a quiet thread is allowed | the moon crater gets two voices and no answer; the strip says「2 voices」and nothing pretends otherwise | filling a thread to a count |
AFTER THE BRIEF IS PUBLISHED (the line's step 9)
10 THE COMPANY 2–4 personas by the match (field × tags × the story's people), picked for different
angles — code + one Flash pick ~0.2¢
11 THE VOICES one call per persona: the brief + the notebook (cached prefix) + its profile + the
「add」rule + the length is yours → the voice + the ONE line it anchors on (must be a
line of the page — lint; a figure outside the brief + notebook — lint; a restatement
of the gist — lint; a throat-clearing opener — lint; one re-ask) 3 × 0.4¢ ≈ 1.2¢
11b THE HEARTS code: each persona in the company hearts the voices that touch its tags, a few from
the roster who share the field; never itself 0¢
12 AN ANSWER only if a voice's line calls for a fact another persona in the company holds
(code asks: does any other voice name this?) → one call, quoting the line rare, 0.4¢
13 PUBLISH brew/<date>/voices/<pid>.json → /ink/voices/<pid>.json · the strip's names on the item
· each voice's rendition in the other language, one call 3 × 0.2¢
ONE BRIEF'S COMPANY ≈ 2¢ · a morning of twelve ≈ 25¢ (off-peak)
The brief line at 3.3¢ and the company at 2¢ make a morning of twelve about 65¢ — under one essay. The persona calls ride the same cached profile prefix the chat room already pays for.
| Date | What |
| 2026-09-29 · the angles | the owner, after echo round two:「have we hit a limit of using an LLM to mimic a human to comment?」— partly: the writing ceiling (the 09-11 loop) is real, but the ECHO is a different failure. Five comments sounded like one machine because every seat answered the same「what would you post?」to the same 300 words and the same three numbers, then read the earlier answers. Three rounds of checks-and-retries each moved one number and grew a new tic (「Newton beat me to it」·「I won't repeat that」) because the judge is the same kind of model that wrote the text. So: give each seat its own angle before it writes. One Flash call a brief (company_angles, ~0.1¢): for each seated persona, from the brief and their card, a passage — one line or figure this person would fix on, never the headline — a stance from a small fixed vocabulary (dispute it · what it costs · who loses · what nobody is asking · why it happened · I've seen this before), spread so no two seats share one, and the ban — the obvious take everyone would make, named once, given to nobody. Not a fixed list of questions (that would be the same five columns under every story): the vocabulary is fixed, the passage and the pairing come from the story. The owner's rulings: the persona still sees the earlier comments (the replies-to-replies stay) and no cap on seats. Measured on Tuesday's 16 briefs, the box's own inputs: |
| | box (echo v1) | echo v2 only | angles |
| later comments mostly repeating one above | 6 | 4 | 1 |
| …repeating some point above | 12 | 11 | 7 |
| word-for-word reuse (5-word runs) | 1 | 4 | 0 |
| over 150 words | 4 | 1 | 0 |
By eye the two worst threads now go three ways: under Starship Musk explains the engine, Feynman asks why the recoverable booster got one line, Liu Cixin prices the capacity curve; under the villas Max Rockbank prices the terrace, Li Ka-shing reads the slow climb in the market figures. ⚠ The first angled run wrote LONGER (nine over 150 words — a steer read as licence); the steer now carries the length itself. ⚠ The judge stays noisy (±1 between identical runs), and the register ceiling is untouched — this removes the shared input, not the model's voice. Kept in each voice record as angle; the brew log names each seat's stance. |
| 2026-09-29 · echoes, round two | the owner:「the repeat / echo problem still exists — check Tue's brew」. The 09-27 checks were live on the box and dropped 13 comments that morning, but reading the edition, sameness still got through: under Starship Liu Cixin and Feynman made the same four points; Li Ka-shing opened with Max Rockbank's line and his Porto Cervo point; Sam Altman's「arguments in the same jacket」and Feddo's「wearing one jacket」. Why: the echo read asked for the main point「with nothing of its own」, so a comment repeating three points and adding one passed; the copy lint compares words, so a Chinese comment echoing an English one was invisible; and over-long comments (533 · 393 words) cover every angle, everyone's included. The fix — the echo read counts points: it lists the new comment's points and marks each one a commenter above already made (agreeing counts as theirs; citing the same fact from the story does not). Half or more taken → the re-ask names them, and a comment still so is dropped; some taken → the re-ask names them and the comment comes back trimmed, never dropped. Two same-named library entries can no longer both sit under one brief. Re-run of the same 16 briefs: mostly-repeats 5 → 3 of the later comments, some-repeats 68% → 52%, over-150-words 4 → 1; by eye the two worst threads now diverge (Feynman asks why the engine quit, Liu Cixin raises refuelling and orbital traffic; Li Ka-shing turns to the upkeep of 48 rooms). ⚠ Honest limits: word-for-word reuse went 1 → 4, the judge is noisy (the same files scored 6 then 5), and a new opener appeared —「I won't repeat that」. If sameness still shows, the next step is structural, not another check: give each seat its own angle before it writes. |
| 2026-09-29 · the reporting behind the brief (local) | the owner, after a chat with Frank Nolan on the SF luxury brief:「much material gathered before this news article is written — a pity Nolan has no access to them」. Short answer: the chat now carries what the brew read, not just the 250 words it kept. Before: the door gave the persona the gist and the three story paragraphs and said「cite nothing outside it」— asked「can you pull data?」, Nolan said he doesn't carry spreadsheets, while the notes held April +43.7%, May +22.7%, 54% of condos over list; and the reader's own question (Lurie's 50 buyers for fewer than 10 $20M homes) came from the page's Q&A, which Nolan never saw. After: the door gives the whole reader page (Q&A, timeline) plus the reporting behind it — the peg's notebook (facts, numbers, people, verbatim quotes, disputed, not established, the cited sources) and the follow-up answers with their sources, ≈2.5–3K words, read once and cached; the rule is「the brief and the reporting are the record; a figure or a deal in it belongs to whoever it names, never to you; where the reporting is silent, say so」(the middle clause after a test turn had Nolan present Lurie's「70% off-market」as「what I move」). The door also reads the brief's OWN day's pegs file now — it read only today's, so an older brief opened with its dek alone. The raw follow-up notes (≈7K words) stay out — the answers already fold them. A forward carries the same block. Measured on dev (Nolan, the 09-22 East Bay brief): every figure he quoted is in the notes, and for Palo Alto/Atherton he said the reporting didn't break them out. One helper, ink_brew.brief_notes_text; an old edition whose brew files are gone opens as before. Existing chats keep what they were seeded with. |
| 2026-09-28 · forward to a chat (v1142) | the owner:「in WeChat a user can forward an article not only to a user but also to a chat」. Short answer: ⋯ › Forward now opens Forward to — the chat-select page message forwarding already uses — with New chat with personas… on top (the v1138 picker) and then your chats, recent first, never re-ranked; a chat where one of the story's commenters already sits says so (Michael Moritz is here). A chat → WeChat's confirm: the card, an optional message, Send. In a chat with personas the panel answers your card and your question like any link you share (measured: 「Is this a real trend or one bad quarter?」→ Moritz:「不是一季的事…」with the brief's numbers); in a people-only chat it is simply your message. One chat per forward (multi-select later); the server's /api/ink/forward takes room_id + text, member-checked, one freeze helper (_ink_fwd_link) for both targets. The chat list and its confirm ride above the article (z97/98) — the v1139 lesson, both layers. |
| 2026-09-28 · forward (v1138) | the owner:「a Forward in the article's ⋯ — open a persona selector, the same as Add persona; pick one or more to forward the article to and start a conversation; rank the more relevant on top」. Short answer: ⋯ › Forward opens the Add-persona page itself (search, sort, friends) under the title Forward to, with the story's people pinned in a Suggested group (a lightbulb heading — the star is Starred's) — its commenters in the thread's order, then the brew's fit ranking (only a real score, ≥ 0.05: the ranking's zero tail is alphabetical), and for an essay its author. Forward makes a new chat with the picks, and the first message is yours — the article as the in-app Ink card (a tap opens it) — which the panel answers like any link you share. The door's mirror: the door has the persona hand you its piece; a forward has you hand it to them. One server call (POST /api/ink/forward) creates, opens quietly (no greeting), freezes the piece as your link and posts it through the same code as a typed message. Found building it: a message with only a link was refused as「empty」and never drawn — a link now counts, like a photo. |
| 2026-09-27 · no echoes | the owner, reading the 09-27 edition:「the later commentor comments on early comments — good — but half of the comments converge, sometimes even say the exact same thing」. Short answer: each persona writes in turn and reads the thread above it (that is what makes the replies work), and the model reused the thread's best lines — Roosevelt repeated Churchill's「that isn't policy, it's a date」word for word; four physicists under one brief each made the same caution. Measured on that edition: 15 of 39 later comments reused a five-word run from an earlier one, 20 of 39 made a point already made. The fix is two lints on the existing silent re-ask, no rule in the prompt: ① the copy lint (voice_copy) — a five-word run, or eight Chinese characters, shared with a comment above, skipping runs the story itself carries (a fact both quote is not a copy); ② the echo read (voice_echo) — one small Flash call on a later comment: is its main point one already made, with nothing of its own? The re-ask names the persona it echoed; a voice still echoing after it is dropped — the thread already holds what it says. Re-run of the same 16 briefs, same inputs: 55 → 41 comments; reused runs 15 → 3 of the later ones (39 → 25); same point 20 → 5. ⚠ The echo count is read by the same kind of call that gates it — the independent measure is the word runs and the owner's read. Watch next: a new opener habit in the later voices (「Newton beat me to the good part, so I'll just add this」), and comments running long after the re-ask. |
| 2026-09-22 · loosened, and in turn | the owner's read of the rewritten comments:「I have no view on whether Anthropic should ship」and「I don't know enough about IPO dynamics」(nobody says what they don't know — they leave it out);「boring lawyer question before believing either headline」(nobody asks the news a question — they dig or shut up); the crater's four from one angle; everyone anchored on 7% —「not more instructions; loosen the prompt so they don't feel they are answering a question」. Traced: 「you may end unsure」+「this is your field」made them cover what they could not speak to; the menu of forms made them perform a question; the numbers list and the parallel run made the anchors and the pile-ups. Now the call is a SITUATION — you are scrolling, this came up, here is what is under it so far, what would you post, if anything — with no forms, no field line, no rules in the text (the lints stay a silent re-ask) and「nothing」as an answer that takes no seat. And the personas comment IN TURN, each seeing the thread so far: a later one may answer an earlier one, take another angle, or pass. The briefs run in parallel, the voices inside each in sequence. |
| 2026-09-22 · the name's door, settled (v1127) | the owner, by screenshot: the「persona detail page」is the PERSONA PAGE — hero, Private message him, the Ink shelf, Profile › Shared chats › Lately › — not the Profile text under it. A comment's name opens that page above the comments (the two Ink pages step under it while it is up). |
| 2026-09-22 · the comment, rewritten (v1126) | the owner:「what we want is the genuine reaction — an angle, a reading between the lines, a punchy comment, an explanation, a disagreement; not a rigid repeat of the story quoting numbers to the decimal; no human comments like this」→ the cause was the call: copy a phrase from the page word for word, use only the record's figures, add from a checklist of four, the page as a JSON dump. Now: the page as prose, the one ask「write what you would actually post under it」with the forms named as choices not a list to cover, numbers the way people say them, no anchor and no underline; the lints keep the figure rule with ROUNDING allowed and add「reads like a report」(three figures to the decimal). The twelve re-run for the owner's compare. |
| 2026-09-22 · three more (v1125) | the owner: names without glyphs on the list row and the Comments bar (the comment posts keep their monogram); the Comments page keeps a ⋯ with Light / Dark; a comment written in another language shows the ORIGINAL with its translation below — Xiao Pang in Chinese, then the English — and no translation when the app's language is the persona's own. |
| 2026-09-22 · the detail page alone (v1124) | the owner:「just open the detail page, don't go to the profile page」→ a comment's name opens the Profile sub-page by itself, titled with the persona's name, from the same cache and fetch the persona page uses; nothing under it. |
| 2026-09-22 · four more (v1123) | the owner: no「Fresh Ink」mast — the date, then the list;「Discuss with him / her / them」from the card's pronoun (the persona page's own pill rule; the short name when the card has none); a name in a comment opens the persona's detail page — the Profile sub-page — with the persona page under it; and yes, every card carries a short name (short: — Moritz, Peak, 小胖), the pickers and the pill use it. |
| 2026-09-22 · two more (v1122) | the owner: in wide mode the story list would not switch stories while the Comments page was open — it did switch, the brief repainted UNDER the open comments; picking a story now closes them. A name tapped in a comment opened the persona page (z88) beneath the brief page (94) and the comments (95); the two step under it while it is open. The name's colour is the persona's own card colour, the one its page and its bylines wear. |
| 2026-09-22 · six UI notes (v1121) | the owner: the company is named the way the chat names a persona — its glyph and its name in its colour — on the list row (no「2 voices」count) and on the brief page as「Comments: Ⓜ Michael Moritz Ⓨ Yokum Taku」, under the date and again after Still unknown; a tap flies in the sub-page, titled Comments; a name there opens the persona page;「Ask X」became「Discuss with X」, a soft pill with a speech mark. |
| 2026-09-22 · k = 5 | the owner, after the cost table (five voices ≈ 0.6¢ a brief, a morning of twelve with five each ≈ 50¢):「set k = 5 and no need for the 0.00 score」→ COMPANY_K = 5, the floors replaced by one rule: a seat needs a number above zero. |
| 2026-09-22 · the line's tags · the window | the owner:「why is Peak Ji low on every AI story — he is a practitioner」→ two findings. The ranker's line had carried a persona's tags ALPHABETICALLY, eight of them: Peak Ji read as「ai, ai agent, beijing, builder, china, chinese, coder …」and manus, frontier models never reached Jev; now the line carries the tags as the author wrote them, the field words first (his tags.txt re-ordered and widened — persona content, committed). Removing the block's「what they are not」sentence made him WORSE (61st), so it stays. And the guessed 26k-token budget drew max_tokens_exceeded once the lines grew: the budget is 18k and a call the API refuses as too big is halved on the API's own word, the halves' tops meeting in a last call. After: Peak Ji 4th–5th on the Anthropic model stories at 0.02–0.06 — a contest Karpathy wins at 0.45–0.54; the company's record now keeps Jev's top eight with their numbers for every brief. |
| 2026-09-22 · no dedupe · more personas | the owner:「remove the dedupe rule — a persona may comment on any brief; clone the box's new personas to dev for testing, never a second copy of one we have」→ the rule out of company_one; the box's app-built roster diffed against both local roots by slug and name, the new ones cloned into personas-built/, their blocks built; the company re-run. |
| 2026-09-22 · the block + Jev | the owner:「fix the roster problem; Jev through OpenRouter; top-k; no outsider, no licence; scrap the proxy; if the input outgrows 32k, divide and conquer; build the block」→ block.txt beside every persona (105 built, ≈3¢), jev_rank with chunking, company_seats = the subject + Jev's top-k on line + block; the bench above; the company re-run on the twelve dev briefs. |
| 2026-09-22 · built on local (v1119) | the owner:「Jev is full — rely on the traditional way and go ahead」→ the company in the brew: company_shortlist (tags weighted by rarity and the field, the subject from the notebook's people, three far candidates) → company_pick (one Flash call, the licence checked word for word against the candidate's line and profile; nobody licensed → the top of the shortlist) → voice_call per persona (its profile the cached prefix, the page + the notebook, the seat's line, the add-don't-restate rule, its own length, cut to the chase; four lints — the anchor is a line of the page · no figure outside the brief and the notebook · no restating the gist · no throat-clearing opener — one re-ask) → voice_rendition (the other language) → hearts_for in code → stage_company writes brew/<date>/company/<pid>.json and /ink/voices/<pid>.json and puts the faces on the edition item; the line runs it after publish; CLI company. The app: the faces and「N voices」on the list item, a strip under the dek, the voices page (the chat's name line, the anchor underlined, hearts with faces, the rendition in the reader's language), Ask → the door with the persona and its post (one room per reader per brief per persona; the seeded page says it is a brief and the post under it). The console: the leaf app.ink_company on Usage & cost and Models & APIs. Owed: a persona answering another (rare, real), the anchor's tap-to-row, readers' hearts. |
| 2026-09-22 · who speaks | the owner:「who should be commenting?」→ §1b: four seats filled four ways — the subject · the expert (a code shortlist, one pick call with the licence) · the outsider · the reader's proxy — with the reader's own personas first, one brief a morning, yield after two days, a quiet brief allowed. |
| 2026-09-22 · the third mockup | the owner's seven notes: no count line under the title · the name line is the chat's (monogram · name in the seat colour · role in grey) · cut to the chase, no throat-clearing openers · each voice its own length, one line to two paragraphs · hearts, and the personas heart each other (code, no call) · they read the brief AND the notebook · the bottom box goes, Ask on every voice is the door. |
| 2026-09-22 · the second mockup | the owner:「for / against is the wrong shape — the persona gives its reaction and thinking, a deeper analysis or a new perspective; interaction optional」and the three goals of Ink → no sides, two to four deeper voices that must ADD, an answer only when real, Ask on every voice. Cost ≈ 2¢ a brief. No code touched. |
| 2026-09-22 · the first mockup | a forum of six takes with stances and a staged argument round; superseded the same day. |