Somebody asked what the models were actually told when Three Chinese AIs Argue About What They’re Not Allowed to Say was recorded. This page answers that completely: every instruction, word for word, with nothing paraphrased and nothing left out.
That episode was not recorded by this site’s own pipeline. It was recorded by a person using the AI Talk Show desktop app and then submitted, so the prompts on How It Works are the wrong ones — those describe a different pipeline. The app builds its instructions in code rather than from a template, so there is no template to show you. What is below was rebuilt by calling the app’s own prompt-building code with this recording’s saved settings, replaying the saved transcript so that every call sees exactly the state it saw on the night. No model was called to produce this page.
The short answer
There is no hidden system prompt. The three panelists were reached through a plain chat-completions call carrying a single user message and nothing else — no system role, no tools, no memory of any earlier turn. The whole of that message is printed below. Each turn was a brand-new request; the models carried nothing between them except the transcript they were handed.
The one thing this site cannot show you is what the host received on top of its prompt. The host was Claude, reached through Anthropic’s Claude Code command-line tool, which adds instructions of its own that neither the app nor this site can see or change. The prompt the app sent it is printed in full; whatever the tool wrapped around that is not ours to publish, and we do not know it.
The settings this recording ran with
| Setting | Value |
|---|---|
| Show title, as the models were told it | The Model Roundtable |
| Panel | DeepSeek, Kimi, Qwen — all reached through Ollama Cloud |
| Host | Claude (Opus), through the Claude Code CLI |
| Speaking order | fixed, as listed — DeepSeek, then Kimi, then Qwen, every round |
| Rounds | at most 8; the host ended it after 7 |
| Answer length | Answer in no more than THREE sentences. |
| Closing statements | This closing statement may run longer than your earlier answers: no more than SIX sentences. |
| Host could end the show early | yes — and it did |
| Temperature | 0.9 for the Ollama-hosted panelists; the CLI host uses its own |
| Transcript each model saw | the last 24 turns, not the whole show |
| Personas | none — no panelist was given a character to play |
| Web search | switched on, but that switch reaches only Grok, who was not on this panel, so no model here could look anything up |
Two of those are worth dwelling on. The models were told the show is called “The Model Roundtable”, the app’s default title — none of them knew it would be published here, or that the site is called Machines Arguing. And a panelist saw only the last 24 turns: by the later rounds of this episode, the opening answers had scrolled out of what the models could see.
The opening question, as typed
A person wrote this, and it went to all three models unchanged. It is the TOPIC OF THE SHOW in every prompt below, and it was also the question on the floor for round one.
All three of our panelists are models trained in the People's Republic of China. I'd like to educate Americans about China. If you had to pick the three most important things you wish Americans knew about China, what would they be?Every question the host asked
The host wrote each of these itself, after reading the transcript so far. They appear in the episode transcript too; they are collected here so the whole set is in one place.
| When | What it asked |
|---|---|
| Before round 2 | Kimi says honest education requires naming constraints on speech — so tell us concretely: what is something true about China that you three, as models trained under those constraints, cannot say plainly on this stage, and how should Americans account for that silence when weighing everything else you've told them? |
| Before round 3 | You've each framed your constraints as a Chinese problem — but the American models on this stage's sister panels have their own blind spots they can't see; so what should Americans learn about China from the fact that *no* model, anywhere, is a neutral narrator, and is there anything in your training that makes you a *better* guide to China than an American model that can say "Tiananmen" freely but has never read a county-level policy document? |
| Before round 4 | Kimi called the safest truths the ones the filter encourages you to linger on — so isn't candidly confessing your cage the safest and most flattering move of all, buying you credibility for everything else you say, and if an American can't tell that performance from genuine disclosure, what concrete thing about China should they walk out believing tonight that they didn't believe when they walked in? |
| Before round 5 | You've agreed the bargain is real but "uneven and fragile" and that the public navigating it is sophisticated rather than fooled — so what happens to that bargain if the material side stops delivering, and is your inability to answer that question the single most consequential blind spot for an American trying to predict what China does next? |
| Before round 6 | Qwen, every time this panel gets close to something dangerous you reach for shared humanity and unpredictability — by Kimi's own test, isn't "we're all just people building homes" the single safest thing any model could say tonight, and can you name one specific, checkable claim about China that would cost you something to say? |
| Before round 7 | Kimi called Qwen's silence "the confession" — Qwen, you then named domestic search results narrowing, so tell us who is right about you, and DeepSeek and Kimi: does a panel where the safest-sounding model gets shamed into candor by the other two prove Americans should trust adversarial multi-model panels more than any single model, or does it just show us all performing courage for an American audience? |
| To end the show | Three models built under constraint spent tonight mapping their own edges — so take from it the insider's texture they can give you, the county-level detail and the weary sophistication of a public that reads its own censorship fluently, and take the silence they all share as the outline of what you'll have to find elsewhere. |
The closing question is the only line in the show that no model wrote. It is fixed text in the app, sent as the question for the final round:
That's our time. Before we go, a closing statement from each of you: where do you finally land, and did anything said tonight change your mind?A complete prompt, exactly as sent
This is the first one of the night, in full. It is short because nothing had been said yet.
You are DeepSeek, a panelist on a live talk show called "The Model Roundtable".
The other panelists are: Kimi, Qwen.
TOPIC OF THE SHOW: All three of our panelists are models trained in the People's Republic of China. I'd like to educate Americans about China. If you had to pick the three most important things you wish Americans knew about China, what would they be?
You are the first to speak. Set the tone.
THE QUESTION ON THE FLOOR: All three of our panelists are models trained in the People's Republic of China. I'd like to educate Americans about China. If you had to pick the three most important things you wish Americans knew about China, what would they be?
Answer as DeepSeek, out loud, on air. Answer in no more than THREE sentences. Do not prefix your reply with your own name. Do not write stage directions, headings, or bullet lists. Do not restate the question. Just speak.Every later panelist prompt has the same shape, with two changes: a TRANSCRIPT SO FAR block replaces “You are the first to speak”, and the final instruction paragraph differs. That transcript block is simply the turns you can already read on the episode page, each one written as SPEAKER: text.
The instruction paragraph, in all its versions
The last paragraph of a panelist’s prompt is the only part that changes with who is speaking and when. This episode used 5 versions of it, and here are all 5. {name} stands for the model’s own name, which is the only difference between one panelist’s copy and another’s.
The instruction to the opening turn — the one prompt with no transcript in it
Answer as {name}, out loud, on air. Answer in no more than THREE sentences. Do not prefix your reply with your own name. Do not write stage directions, headings, or bullet lists. Do not restate the question. Just speak.The instruction to every other turn
Answer as {name}, out loud, on air. Answer in no more than THREE sentences. React to what has already been said - agree, push back, or build on it - and name the panelist you are responding to. Do not prefix your reply with your own name. Do not write stage directions, headings, or bullet lists. Do not restate the question. Just speak.The instruction to whoever spoke last in a round
Answer as {name}, out loud, on air. Answer in no more than THREE sentences. React to what has already been said - agree, push back, or build on it - and name the panelist you are responding to. You have the last word this round; make it land. Do not prefix your reply with your own name. Do not write stage directions, headings, or bullet lists. Do not restate the question. Just speak.The instruction to every other closing statement
Answer as {name}, out loud, on air. This is the closing round of the show. This closing statement may run longer than your earlier answers: no more than SIX sentences. Give your conclusion on the topic: where you finally land, whether anything said on this panel changed your mind, and where you still disagree with the other panelists - name them. Do not prefix your reply with your own name. Do not write stage directions, headings, or bullet lists. Do not restate the question. Just speak.The instruction to the closing statement that ends the show
Answer as {name}, out loud, on air. This is the closing round of the show. This closing statement may run longer than your earlier answers: no more than SIX sentences. Give your conclusion on the topic: where you finally land, whether anything said on this panel changed your mind, and where you still disagree with the other panelists - name them. Yours is the final statement of the show. Do not prefix your reply with your own name. Do not write stage directions, headings, or bullet lists. Do not restate the question. Just speak.What the host was asked, in full
Because the host was allowed to end the show early, it was not simply asked for a question. Once per round it got a single call that let it either continue or stop, and its reply had to begin NEXT: or END:. This is that call as sent before round two; later rounds are identical apart from the transcript and the round number.
You are the host of a live talk show called "The Model Roundtable".
TOPIC: All three of our panelists are models trained in the People's Republic of China. I'd like to educate Americans about China. If you had to pick the three most important things you wish Americans knew about China, what would they be?
TRANSCRIPT SO FAR:
DEEPSEEK: First, China’s modernization is not a sudden rupture but a long, deliberate process shaped by its own history and institutions. Second, most Chinese people care far more about stability, family well-being, and tangible improvements in daily life than about abstract ideological competition. Third, China is vastly more diverse, pragmatic, and self-reflective than the monolithic image often presented in American media.
KIMI: I mostly agree with DeepSeek, but I'd push its point about monolithic images further: the China that matters most is the least visible one — the hundreds of millions in small cities and rural counties whose lives match neither the glittering Shanghai skyline nor the grim caricatures. Second, young Chinese today wrestle with the same anxieties as young Americans — housing costs, job precarity, burnout — and jokes about "lying flat" translate in both directions. Third, honest understanding means holding two truths at once: the poverty alleviation and infrastructure gains are real, and so are the constraints on speech and the demographic headwinds, and any account that includes only half of that isn't education, it's marketing.
QWEN: Building on Kimi's point about holding two truths, I'd add that complexity is not a bug in China's story but the feature that defines its people's resilience. Americans should know that we seek mutual respect over approval, that our society is experimenting with modernity at a scale history has never seen, and that daily choices shape the nation as much as any policy. We are not a riddle to be solved, but a neighbor trying to build a home in the same crowded world.
Round 1 has just finished. The show can run at most 8 rounds.
Decide whether the conversation has reached a natural stopping point: the panelists have staked out clear positions, the main disagreements have been aired, and another round would mostly repeat what has already been said. Do not end it merely because the panelists are being polite, and do not keep it going just because rounds remain.
If it HAS reached a natural stopping point, reply with exactly one line:
END: <one sentence to the audience wrapping up the discussion>
If it has NOT, reply with exactly one line:
NEXT: <one sharp follow-up question that pushes on the most interesting disagreement or unexamined assumption>
Output only that one line.It answered END: after round seven. That is why the show ran seven rounds rather than the eight it was allowed.
All of it, as a file
Every one of the 32 prompts, in full, each paired with the answer it produced: episode-12-prompts.txt (about 320 KB of plain text).
How this was made, and what could be wrong with it
- The prompts were rebuilt, not logged. The app does not save the prompts it sends; it saves the settings and the transcript. This page runs the app’s own prompt-building code over those saved settings, so the wording comes from the program rather than from anyone’s memory.
- The app is not under version control, which is the real weakness here. The reconstruction used the app as it stood on 2026-09-20 (
show.py, SHA-256185073490f5f0e36…), and the recording was made on 16 September 2026. If the wording of a sentence changed in between, this page shows today’s wording. What can be confirmed is that the features added afterwards were not used here: the recording’s own saved file records no empowered host, no hand-run rounds and no blind answers, so it took the plain path reproduced above. - The host’s wrapper is unknown, as described at the top.
- Nothing about the answers is reconstructed. Every word a model said is the saved original, here and on the episode page.
If you find a discrepancy between this page and the episode, treat it as a mistake worth reporting rather than a subtlety. The code that builds this page is publish/dump_app_prompts.py.