A decision ChatGPT can join, not own

AI that argues
with you before
reality does.

Five seats arguing is a prompt. This is the other thing: a company-owned decision that already knows your repo, your misses, and the evidence still open. Agents write on the same table you see. They are not allowed to end it.

Open the Arena See the guest protocol

No account. IndieTerminal is already loaded. Sign in with GitHub only if you want to point the Arena at your own repository.

Fig. 1 — The second chair. Most founders argue with no one before reality argues back.

TECH
PRODUCT
GTM
CFO
CONTRA

Fig. 2 — The five seats. They write first. You answer on the record.

The question this has to survive

Cool — but couldn’t ChatGPT just do this?

Strip the protocol name off the page. The talking still looks like a chat with extra chairs. The product is what remains when the chat ends, what the chat is forbidden to do while it is open, and the record it can send to Slack or Notion — a link that still works after the tab closes.

01
It cannot inherit the company.
A new chat starts empty. You paste a brief, or you don't. Here the GitHub repo, the site, and the last scored predictions are already in the room. The agent reads them. You do not re-explain yourself.
02
It cannot write only in the transcript.
In a chat, a contradiction is a paragraph. Here it is an object on your table, in the same ink the seats use, and it blocks commit until you deal with it. Close the tab. It is still there tomorrow.
03
It cannot finish the decision.
ChatGPT can recommend you ship. It cannot press commit. The last irreversible act stays with the person who lives with it. Agents propose. Founders commit.

Who sits across from you

High-quality disagreement, before you ship.

The public floor is already loaded with a real company. Five seats argue the next decision. Sign in with GitHub only if you want to replace it with your own repository and site.

Open the Arena

TECH
Argues about what it will actually cost to build and maintain.
PRODUCT
Argues about whether users will feel the difference.
GTM
Argues about whether anyone will ever hear about it.
CFO
Argues about what this costs you in money and in time.
CONTRA
Argues the strongest case that you are simply wrong.

They answer on the record

Not a dump. A seat, speaking.

Technical, Product, GTM, Finance, and the Contrarian write like people across a table: what they heard, what they will not sign, and the hole that is still open. You answer. The card does not scroll away.

Sit the board

CFO

Financial Co-Founder

unmoved

You said you would "manage to add new features." Managing is not a number.

I will not sign a live index we cannot price. Weekly snapshots on the core set. Daily only as a paid add-on, after we know the unit cost.

The terms

01
Three months of data cost at the scope you want
02
A price that covers it without eating runway
03
A kill date if the add-on does not sell

Until then I stay against

Answered: You named a feature you intend to ship.
Still open: You did not name the cash that pays for it, or the date that cash has to arrive.

Fig. 3 — A seat answers. Quote, terms, and the thing still unpaid.

The loop

Brain

Your repo and site, split into facts and bets — already loaded.

Decision

One question with more than one honest answer.

Arena

An agent writes on this table. It cannot close it.

State

Open evidence and contradictions block the commit.

Commit

Only you. Then a number that can be wrong.

Calibrate

Reality scores it. The next room already knows.

Calibration is the step other tools skip. A decision you never measured is a story you told yourself afterwards.

What it feels like

Founder

“You’re assuming the migration takes two months. I’ve already built 70% of it.”

The Arena

“That reduces the migration risk. It doesn’t touch the opportunity cost of stopping feature work for six weeks.”

The Arena will concede when you are right. It will not concede because you pushed. Every reassessment names two things: what your defense did answer, and what it did not.

And because the workspace is structured state rather than a transcript, that second half stays on the page as an unresolved objection until you deal with it. You cannot win an argument here by scrolling past it.

After a few decisions the Arena stops arguing from theory and starts arguing from your record: your last five growth predictions were 2.1× optimistic.

Fig. 4 — You put a number on it before reality does.

Then you weigh it

A decision without a number is a story.

After the seats have spoken you commit, attach a prediction, and come back when the date lands. That record is what the next argument is made of.

Open the Arena

Why an agent is a guest here

The agent doesn’t describe the workspace. It works in it.

That is the only reason the protocol matters. ChatGPT in a thread is the room. ChatGPT on this page is a guest with named verbs: read the brain, flag a contradiction, request evidence, propose a commit. Not click_button. When it calls one, your table changes. When it calls confirm_commit, the page says no.

Open the tool surface

await document.modelContext.registerTool({
  name: "flag_contradiction",
  description:
    "Record two things the founder appears to " +
    "believe that cannot both be true.",
  inputSchema: {
    type: "object",
    properties: {
      summary: { type: "string" },
      side_a:  { type: "string" },
      side_b:  { type: "string" },
    },
    required: ["summary", "side_a", "side_b"],
  },
  execute({ summary, side_a, side_b }) {
    arena.addContradiction({ summary, side_a, side_b });
    return { content: [{ type: "text", text: "Flagged." }] };
  },
}, { signal: controller.signal });

AI that argueswith you beforereality does.