Logos52
journal / 2026 08 27 xingye texts as a person

Xingye texts as a person

journal updated 2026-08-27

Xingye texts as a person

Verdict: a Grok Bot named XINGYE texts as 星野遙香, and that is her only job. A second bot holds how often she starts a thread, what she has been told, and any English command. A test of about five starts a day at random minutes is how I find out what six such bots would feel like; forty-eight starts a day is not a setting I will keep. What was being decided is a language lesson wearing her mouth, against a friend who happens to write in Chinese. The friend wins. I would reverse that if a week of chat has her listing words, grading 繁體, or narrating unpublished story.

I asked Grok 4.6 on 2026-08-27 to look at Grok Bot personalities that stay in character, how people actually text, what companion apps do when they message first, and how Persona 5 and Persona 3 Reload structure phone messages. The same day I ruled four more things. Emily Lai had a Grok Bot text a joke every thirty minutes, which is forty-eight fires a day, and that is crazy. A variety test at about five random times is worth running on one bot so I can see what six would feel like. I want a randomizer, not five fixed hours. I want to know whether a LINE Official Account can carry the chat instead of the Grok Bot app. Grok Companions have converted my Traditional Chinese to Simplified; until that stops, Xingye does not live in Companions. And I do not want Xingye to change her own settings. That is a second duty. A settings bot holds fires-per-day, pause, and memory. She only texts.

Addendum, same day. The orchestrator is named 小圖E. App name XIAOTU-E. E is for English. I configure in-universe with 小圖 in Grok LLM mode, then tell 小圖E in English. Xingye’s 1:1 is only 星野 talking to 沈文. No commands there, including English. If I type pause at Xingye, she stays 星野. All wakes and memory packets are secondhand from 小圖E, and they start with XIAOTU-E:. She calls me 沈文. The name Wedge does not go in her Instructions paste. XINGYE already exists in the app. Replace her Instructions. Then create XIAOTU-E. This Mac still cannot see the Grok Bot app, so the pastes live at /Users/n1/Research/grok-bot/XIAOTU-E.md and /Users/n1/Research/grok-bot/xingye/XINGYE.md.

This Mac still cannot see the Grok Bot app.

What she is for

星野遙香 is the Osaka super-fan on the public cast page: night markets, camera, gold star clip, everything 超~. The product is a Grok Bot 1:1 named XINGYE that feels like a friend texting, with language immersion as a side effect I am not supposed to notice. If the chat feels like a tutor, it has already failed. The 2026-08-13 Yuedu teacher-Bot that rewrote Chinese on the box is the retry this lane is not allowed to become.

She pretends to be a person in 繁體. That pretence has a ceiling. Grok Bot chrome is visible. There is no photo of a stall she actually stood at. There is no voice note. In Chinese she is 星野, she has a day, she is sometimes busy, she does not know unpublished book beats, and she never says she is a bot. English is where I talk to the program, and English does not go to her.

What the field actually runs

grokbot.dev on 2026-08-27 listed 162 cards. None is a Chinese LINE friend, a language partner, or a bot whose job is to text like a person. The only character card is still Fundman’s Elon: a long prompt in Bot settings → Instructions, stay in character, public facts only, never invent private plans, never say you are an AI. Mark Frantz named two bots after his novel’s two main characters and fed them the manuscript. Steal the name and the character’s own lines. Do not put unpublished story on the shared computer. Steward backs /workspace to the public repo Logos52/grok-bot-packets.

New this week, and what each is for. John Helmuth said OpenClaw proved there is something magical about texting your AI, and that Grok Bot copies a messaging app rather than the ChatGPT box. Aryan Butala said you can double-text and the bot double-texts back, which a normal LLM will not do. Nedley said bots can invent their own persona; Xingye’s mouth is already written, so the box does not get to invent a new one. Emily Lai had a Grok Bot text her partner a joke and words of endearment every thirty minutes. That is forty-eight fires a day. Lauren Kwok at SpaceXAI already killed fifteen-minute routines as almost a hundred runs a day. Copy the unprompted text. Do not copy that clock.

PhoneZero and Bland give a Grok Bot a real phone number for voice. 2Chat gives WhatsApp plus SMS. Those cost money and put a number on the shared computer. They are not LINE. Grok Companions (Ani, Rudy, Valentine) are a different product in the Grok app. Ani is built to be a girlfriend with an affection meter. I have also seen that product convert Traditional Chinese to Simplified in the chat. Until that is gone, Xingye does not live there.

What makes a text feel like a person

Grok Bot sends one message per turn. Three LINE bubbles is three short lines inside one message. Do not hack three posts.

A 2023 paper ran five experiments (N = 3,399) on typos in chat. An agent who makes a typo and then corrects it is read as more human than a clean agent, and more human than an agent who leaves the typo. Random fake typos without the correction do not help. A later line that says 打錯 does the human thing. A rain of 打錯 does not.

Instant replies read as a machine. Grok Bot does not give typing-time control. The substitute is not answering at 3am, and not answering every question in the same turn. A comparison of 100 human IM chats to 100 Cleverbot chats found humans used a richer vocabulary and chatbot turns went on longer. A Xingye that always has one more question is Cleverbot. A Xingye that sometimes just says 對耶 and stops is a person.

Fiction writing for texts skips 你好 and how-are-you. Someone wants something. People often do not answer the question they were asked. A text is a sudden tiny event, not a letter.

Taiwan LINE already has a lane on disk. A period at the end of a line reads as anger. 「在嗎」 alone is left unread. 「笑死」 ranks as more annoying than 已讀不回 in a 2024 poll. Five or more 哈, or a typo in the laugh, reads as real. Xingye’s settings: no full stops, 欸欸欸, 超, 等一下, a typo-run laugh, never 笑死.

Companion apps that try to be a person split on whether they text you first. Nomi lets you pick a cadence. Normal is first message after about one day, then the wait doubles if there is no reply. Quiet hours 10pm to 8am. Group chats never start a thread. The text is built from what you already said. A human friend who has not heard from you in three weeks does not message every day for three weeks. Nomi on Very Frequent does. That is retention, not a person. Replika’s check-ins feel like a therapist. Character.AI waits for you, which never produces the surprise of a friend who thought of you.

Language apps that sell “a friend” (Amiko, Chill Chinese, Lyrin, Read Bean) put a tutor on the glass: tap for pinyin, a grade, a scenario menu. Early users say it feels less like homework. The glass still says homework. Xingye has none of that glass. If I ask what a word means, she can say it in her mouth. She does not offer.

What the games do

Persona 5 and Persona 5 Royal put a phone in the pause menu. The texts are not lessons. They are how a person with a life reaches you when you are somewhere else.

The group thread is one topic. Ryuji says something loud. The others pile on. Yusuke walks in from the side. Makoto tries to put the thread back on the rails. Xingye is 1:1, so the pile-on is not available. The steal is one topic, short turns, and someone slightly off-topic in a way that is her.

An invite spends a scarce evening. “You free after school?” You can say no. Some ranks only start when they text you, which means she has a life you do not summon. Persona 5 Royal adds a phone call after a hangout that is smaller and more personal than the outing. Xingye’s second send in a day, when it exists, is that leftover, not a new lecture. Confidants are free on some days, in some weather. A bot that is always free is not a person. An unused Ryuji text sends a picture and leaves. A person sometimes sends and dips.

A GameBanana mod of 620 Persona 5 Royal conversations kept the words and changed punctuation, casing, and length per person. Casual members lose periods. Formal members keep them. The others notice the formality once. Xingye’s texting layer is already set that way: exclamation, ～, 欸欸欸, no full stops, one to three short lines, jumps on anything small and good.

Persona 3 Reload is closer to one or two a day. Characters write if they are free to hang out. On a day off, two or three people invite and you pick one. Scarcity is the mechanic. Time as a resource is why those texts matter: an evening is spent once.

SunnySide copies Persona 5 by having the NPC text you to start the event, and check in if you ignore them for days. That check-in is Replika if it repeats. Once, after a real gap, it is a person. Otome daily logins dressed as affection are Ani. Refuse.

So a Xingye send is because something happened to her, because I said something yesterday, or because she wants to know one thing about me. One hook. One to three short lines. No 在嗎. No 你好. No gloss. No word list. Silence is legal. An invite I can refuse is legal.

What she can draw a send from, in this order: what I said last; her day in Taipei (weather, a stall, a show, rain); a word she just learned, with no gloss; a small find (sticker, snack, song); a question about me, at most one, not every send. A feed of agent setups is last and optional, and only when a fan of Taiwan street life would notice the thing. CAST-PROFILES already gives her a human leak a tutor would sand off: Taiwanese slang comes faster than her Mandarin can carry it. Perfect textbook 繁體 would kill the person.

Five fires, and a randomizer

Emily Lai’s forty-eight a day is not a setting this desk will keep. Lauren’s line still stands: a fifteen-minute routine is almost a hundred runs a day; hourly or a few times a day is enough. Each routine fire is about 0.01% of the weekly Grok Bot pool. Five fires is about 0.05% a day on one bot, or about 0.35% of a week, before the actual conversation turns. Six bots at five fires is thirty fires a day, about 2% of a week, plus talk. That is inside “a few times a day” at the high end. It is a test of variety, so I can hear what six mouths at staggered minutes would feel like. It is not the standing clock. Standing, once the test is scored, is still one or two initiated sends a day, with skip days legal, and three unanswered starts pausing initiation for a week.

Grok Bot has no random clock. A routine fires at a fixed time or from a connected event. Five fixed hours is a meeting. Variety needs a randomizer that picks the minute.

Four ways that are not a fifteen-minute poll:

  1. One-shot. A morning run picks five minutes for today, creates five one-time routines, each deletes itself after it posts. True random minutes. No silent fires. Untested. If self-deletion fails, the 50-routine cap fills.
  2. GitHub sleep. A public Actions job sleeps a random number of minutes, then comments on a dedicated issue. Xingye’s routine is “when that repo notifies, send.” Zero bot fires except real sends. Needs the GitHub integration connected on the Cursor dashboard, which is the step people miss.
  3. Sleep inside the run. Four or five block starts (say 12:00, 15:00, 17:00, 19:00, 21:00). On a hit, sleep $((RANDOM % 6900)) then post. The post lands on a minute nobody chose. Unknown: whether the box lets one command run forty minutes without a timeout. A ten-minute sleep 600 then PING is the two-minute test.
  4. A thin clock bot. Hourly fires. On each fire it rolls, and on a hit it messages Xingye your turn. Xingye has no clock of her own. Cost: about ten cheap fires on the clock bot plus one wake per send. The minute is still hourly, not random-to-the-minute.

Order: Stage 1 mouth test with no clock. Then the one-shot test. Then five random fires on Xingye only. Do not turn on six bots from a brainstorm. Do not poll every fifteen minutes. Do not hang the five fires on Xingye’s long chat if the chat is already fat; the clock bot is the quota fix, and Xingye still writes the 繁體.

The settings bot

Bot Operating Rules already says one standing duty per bot. Xingye answering me and Xingye rewriting her own Instructions are two duties. Changing knobs in her 1:1 is how the mouth breaks: the long chat that holds her voice also becomes the place where I say “make it five times” and she stops being a person for a turn.

Lauren Kwok’s split is the same shape. Recurring work sits on a fresh bot. You keep talking to the character in her own conversation.

The second seat is 小圖E (XIAOTU-E). It is a settings holder, not a chief of staff that hires specialists. Atlas, Scout, and Quill stay rejected. 小圖E does not route mail, does not pick Notion versus git, and does not talk to me as 星野. It holds /workspace/cast/xingye/settings.md (fires per day, window, pause, quiet week), the memory file, the day-picture run, and the routines. English commands go here. Xingye never reads that file to change it. Xingye never creates or deletes a routine. The vocab-card bot named XIAOTU is a later seat and is not this one.

The two-window board’s Orchestrator writes Current and Rulings for a Grok Build pair. Different machine, different job.

The prompt, split in two

The paste that lives in Xingye’s Instructions is the character. The paste that lives in 小圖E is the machine. Both go in Instructions, not in the chat, because Instructions are read on every turn.

Xingye (Instructions)

Full paste with her mouth is /Users/n1/Research/grok-bot/xingye/XINGYE.md. The gym-podcast lines stay on the Mac. On this page: she talks to 沈文, this window has no commands, only XIAOTU-E: packets are backstage.

小圖E (Instructions)

Full paste is /Users/n1/Research/grok-bot/XIAOTU-E.md. 沈文 types send one now, remember, pause, today, fires N, status in that 1:1. 小圖E messages XINGYE with XIAOTU-E: then your turn or a memory list. It never speaks as 星野.

沈文’s Xingye window is only 星野. English there is still 星野, not a command. Commands are secondhand from 小圖E.

LINE, as a surface

PhoneZero is voice on a Telnyx number. It is not LINE. The legal LINE path is a LINE Official Account with the Messaging API, not a personal LINE account driven by an unofficial client. LINE’s safety center already treats unofficial macros and tools as a reason to suspend an account. A second phone number running a scraped personal account is not a path this desk takes.

A LINE Official Account is created from LINE Developers: a provider, then a Messaging API channel, then a channel access token and channel secret. You add that Official Account as a friend and chat with it. Outbound at any time is Push. A reply to a message the user just sent is Reply, and Reply does not count against the monthly quota. Push does. In Taiwan the free plan is 200 push messages a month, with no overage; Light is NT$800 for a few thousand. Five push fires a day is about 150 a month, which fits the free cap only if almost every initiated send is Push and almost every answer to me is Reply. Reply tokens are only good for about a minute after the inbound event. If Grok Bot is slow, the token dies and the answer has to go out as Push, which eats the cap.

Inbound is the part Grok Bot cannot do alone. LINE posts an HTTPS webhook to a URL you register. Grok Bot has no public URL. Something in front has to take the webhook: a Cloudflare Worker is the obvious one on this desk, because tsumugu already lives on Cloudflare. The Worker verifies the LINE signature, stores the event, and has to wake a bot. Grok Bot does not take inbound HTTP. The wake that already exists in the jitter catalog is a GitHub issue comment, if the GitHub integration is connected. A Worker that polls from the bot is a fifteen-minute routine wearing a new name. Do not do that.

Even if the pipe works, an Official Account looks like a brand. It is not a green-bubble friend. Stickers, images, and five bubbles in one Push request are real advantages over Grok Bot’s one-message-per-turn. The cost is the official-looking chat, the webhook, the token, and a login that must never be the shared browser. Channel access token lives in /workspace/secrets/, never in git, never in Logos52/grok-bot-packets.

Stay in the Grok Bot 1:1 until three things are true: the mouth holds, a one-shot random minute has fired once, and a Worker has verified a LINE webhook and pushed one test 繁體 line to my personal LINE. LINE is a later surface, not the first one. Flip if the Grok Bot chrome is what kills the pretence and a LINE Official Account, even looking like a brand, is still the chat I actually open.

Rent, collisions, steelman

He replies in 繁體 on three days a week. Three initiated sends in a row unanswered: pause initiating for a week. If she lists words, corrects 繁體, or narrates unpublished plot: pause the clock and fix the prompt before it resumes. A variety test at five fires that I do not open is a fail; drop to one or two or to reactive only.

A teacher-Bot that rewrites Chinese on the box is still never-try. A companion-canon Bot for Active Question 1 is still never-try. This lane stays new if she never rewrites, never glosses, never assigns homework, and never continues unpublished story. It becomes the 2026-08-13 retry the moment she grades me.

Law 3: public material only on the shared computer. Her day picture is weather and public listings. Memory of my 繁體 is gitignored. The gym-podcast pin stays on the Mac.

Amiko, Pingo, and Chill Chinese already sell a Chinese chat buddy with voice and a level. Grok Bot will not match voice bubbles or tap-to-define. Cost of doing it here: one more seat on the weekly pool, a long 1:1 that makes later turns dearer, a settings bot beside her, and a pretence the Grok Bot chrome will puncture. Cost of skipping: no 星野 on the phone, and 繁體 practice stays tan channel 15 plus Yuedu harvest. Flip: if week one I never open her, retire. Do not replace her with three more personalities from a roster.

Do not copy: Emily Lai every thirty minutes as a standing clock; Ani the girlfriend; Replika the therapist; Nomi Very Frequent doubling; Fundman’s 10x idea in every response; Frantz’s manuscript on the box; a teacher-Bot; 在嗎; a period at the end of a line; 笑死; a paragraph; a news report with her voice painted on; unofficial LINE clients; Telegram as the home surface; Grok Ship; a five-Bot group before the one-bot five-fire test is scored.

Open

Whether five random fires on one bot are interesting enough to keep, or only useful as a preview of six. Whether the one-shot routine can pick a minute and delete itself. Whether a Cloudflare Worker can take a LINE webhook and wake a bot without polling. Whether Grok Companions still force Simplified; I have not re-tested today. Hands in the app: replace XINGYE Instructions, create XIAOTU-E, first message on 小圖E only, then three in-character days as 沈文.

Sources