Support

How to get help

Write to [TO CONFIRM: the address people write to for help]. It is an inbox rather than a queue: there is no chat widget, no ticket number and no status page, and nothing in the product opens a case on your behalf.

Say what you were doing and what the screen said. You do not need to send us your audit reply, and you should not: what you paste into the product is never stored, and an email is the one place that sentence does not reach — it sits in an ordinary inbox like any other message.

No response time is promised here, because none is promised anywhere else on the site either.

“That is the prompt, not the answer”

What you pasted matched several distinctive lines of the audit prompt itself, so the reading stopped rather than treating the questions as answers. It is the ordinary mistake: the prompt is the thing you just copied.

Send the prompt to your assistant, wait for it to write back, and paste that — the whole reply, all six sections and the score it ends on, rather than a summary of it. Nothing from the first paste was kept, so there is nothing to undo and re-pasting costs you only the ten seconds.

Your sign-in link says it has been used

Links work exactly once and expire 30 minutes after they are sent. Opening one twice — or letting a mail client follow it for you — spends it. Ask for another on your settings page; the new one comes back to that same page.

Nothing published is affected by this. Your context, the versions behind it, the connections that hold a credential and every receipt are all still there; the link is how you get back in, not what you own.

Two things worth knowing about that form. It answers identically whether or not the address is known to us, so “check that inbox” is not a confirmation that an account exists — that is not something a form should be able to tell a stranger. And there is no password and no recovery question: if you no longer have the inbox you published to, there is no other way in.

You published, and nothing was attached to you

That is the flow working rather than failing. Publishing writes your context immediately and unowned, because the alternative is losing the decisions you just made to a sign-up form. Typing an address is not proof of holding it, so ownership arrives only when you open the link sent to it.

Until then the context is reachable from the browser that published it and nowhere else: it will not appear on your other devices, no AI can be connected to it, and it is deleted 30 days after publishing if nobody claims it.

Open the link, come back to the tab you published from, and press I have opened the link. If the mail never arrived, ask for another from your settings page using the same address you published with.

ChatGPT is not using your context

A connection is a tool the assistant can call, not something that runs on every message. Its instructions tell it to fetch before answering anything whose quality depends on your current roles, goals, priorities, projects, preferences or constraints — but the model decides, and models under-call tools far more often than they over-call them.

The reliable signal is naming the connection in your message: @Ctrl in ChatGPT, picked from the list that appears. That turns a judgment call into an instruction.

Then check whether a fetch happened at all. Every disclosure writes a receipt before anything is handed over, so the receipts on your settings page settle the question: a row means your context was read, and no row means the answer you got came from the assistant’s own memory. Two things make a fetch impossible rather than unlikely — a context nobody has claimed yet, and a credential that has been revoked.

Stopping an AI from reading it

Revoke the connection on your settings page. Each AI holds its own credential, so revoking one revokes that connection and no other, and it takes effect immediately: the next fetch on that credential fails. Connecting the same assistant again mints a different credential rather than reviving the old one.

It cannot reach backwards. Context already handed to an assistant is in that provider’s conversation and its logs under that provider’s terms, and nothing here retrieves or deletes it. Revoking stops the next fetch; it does not unsay what was already said.

Revoking is also not deleting. The credential’s row is kept and marked revoked, so your old receipts can still name the connection that fetched what.

Deleting everything

There is no self-serve delete button today, for a context or for an account. That is the state of the product rather than a policy, and saying otherwise would be the easiest sentence on this page to write and the first one you could catch us on.

Ask, and we will delete the rows: [TO CONFIRM: the address people write to for deletion and access requests]. That is the same address the privacy page names, and what is stored to delete is the list in section four of it.

One thing deletes itself: a context nobody has claimed is removed 30 days after it was published, without anyone asking.

The rest of it

• Your settings — your context in full, every version behind it, which AIs hold a credential, every fetch they have made, and the revoke button.
• Privacy — what is processed, what is stored table by table, who else handles it, and what the encryption does not promise.
• Terms — what the service does, what it explicitly does not do, and the terms you use it under.