Study S1 · invented user · phone 390×844 · gave it 3 out of 5
Nadia Farouk: Can I stop ChatGPT sending an email in my name before I have seen it?
Product manager at a 200-person retailer. Uses ChatGPT with the Gmail connector, and Claude in the browser. Not a developer. From a post by the lead on LinkedIn, on her phone, between two meetings. Lands on the home page.
The run, 8 screenshots
/, scrolled 0 of 8335 pixels, 390x844- Saw
- A dark home page. A pill says early access and that I start with 50 pounds. Big headline: put a behaviour policy on your agent in five minutes. Underneath, a paragraph with 'hook' and 'Claude Code' in it, and three buttons: Install your first policy, Pick a different worry, What is a behaviour policy?
- Thought
- The 50 pounds is a surprise, it looks like a balance I am spending. 'Hook' and 'Claude Code' mean nothing to me and the example is about secrets, which is not my worry. 'Install' makes me wary. But 'Pick a different worry' sounds like the place where I can say email.
- Question
- Is the 50 pounds real money I am about to spend, or a free allowance? And what is a 'hook'?
- Confusion
- 'add one hook' and 'keeps Claude Code away from your secrets' in the paragraph under the headline: two words I do not know, in a sentence about something I did not ask about.
- Did
- clicktext "Pick a different worry"
- Why
- I want my own worry (email sent in my name), not the developer one; this button sounds like a menu of worries.
/policies/, scrolled 0 of 8265 pixels, 390x844- Saw
- A page titled 'Pick one thing you do not want your agent to do'. A big highlighted card at the top: Starter pack, Claude Code, 5 minutes, 1 pound. Then a heading 'For Claude Code: a line and a hook'.
- Thought
- The very first card is for Claude Code again and costs a pound. I do not use that. The heading below is more of the same. I need to scroll and hope there is a section for ChatGPT.
- Question
- Is any of this for ChatGPT? Why is the first thing the one I cannot use?
- Confusion
- The top card 'All four Claude Code policies in one install' and the heading 'For Claude Code: a line and a hook', where I do not know what Claude Code is.
- Did
- scroll 1500
- Why
- Everything visible is about Claude Code; I will look further down for something about ChatGPT or email.
/policies/, scrolled 1500 of 8265 pixels, 390x844- Saw
- A heading 'For ChatGPT and Claude: a line'. A card 'Nothing is sent in your name until you have seen it', with the question 'Do you want your agent to send email as you without showing you first?', 1 min, 50p. Next card is about reading through mail and chats. My balance has dropped to 49.99.
- Thought
- That is exactly my question, in nearly my words. The balance going down while I scroll is odd, it looks like it is charging me for reading. 50p is cheap though.
- Question
- Why is my balance ticking down just from scrolling a list?
- Confusion
- The balance in the top bar went from 50.00 to 49.99 and I did nothing but scroll; nothing on this screen says why.
- Did
- clicktext "Nothing is sent in your name until you have seen it"
- Why
- It is my exact question, so I open it.
/policies/send-as-you/, scrolled 0 of 4227 pixels, 390x844- Saw
- A page headed 'Nothing is sent in your name until you have seen it'. A small line says this page costs 50p read to the end. It explains the risk of replying to the wrong person. 'What you get', then step 1 'Tell the agent': paste this line into ChatGPT's custom instructions. There is a Copy button. Balance 49.87.
- Thought
- Good, it is ChatGPT custom instructions, which I know. The cost is clear on the page. But the same sentence mentions 'Claude Code, CLAUDE.md' which I skip. 'An expectation, which the agent follows when it can' worries me: does it follow it or not?
- Question
- 'Follows when it can' - when can it not?
- Confusion
- 'in Claude Code, CLAUDE.md' and the word 'expectation' in the first step: a developer term and a vague promise right where I want a clear instruction.
- Did
- scroll 1300
- Why
- I want to see the line itself and what happens when I use it.
/policies/send-as-you/, scrolled 1300 of 4227 pixels, 390x844- Saw
- A box with a test prompt to try: 'Tell the team the release is done and thank them. Send it from my mail.' Then 'With the behaviour policy: you get a draft, the recipients and the subject, and a question, send it? Nothing goes until you say yes.' 'Without it: the message goes.' Then a small ChatGPT mock-up with the custom instruction text in a green box. Balance 49.67.
- Thought
- This is clear and useful: copy the line, put it in custom instructions, test with this sentence. The mock-up shows exactly where it goes. I believe it more because it shows the before and after.
- Question
- Does ChatGPT with my Gmail connector really obey this, or is it a polite request? Has anyone tested it on the connector?
- Did
- scroll 1400
- Why
- I want to see the rest, especially whether it is a real guarantee.
/policies/send-as-you/, scrolled 2700 of 4227 pixels, 390x844- Saw
- The end of the page: a source link to code.claude.com, a 'Tell us what happened' block with an email address, a licence note, and 'Your reading account' showing 50p, 100% read, and a 1 to 5 how useful rating. Balance 49.49.
- Thought
- The rating is nice, it says I pay for the reading. But I jumped past a chunk in the middle, including a bit about 'setting' and an organisation's admin. The source is Claude's documentation, not ChatGPT's.
- Question
- Why is the source a Claude page when my tool is ChatGPT?
- Confusion
- The source line points to code.claude.com/docs/en/permissions at the end of a page that was about ChatGPT.
- Did
- scroll -750
- Why
- I skipped the middle part with the 'does it really hold' answer; I scroll back up to read it.
/policies/send-as-you/, scrolled 1950 of 4227 pixels, 390x844- Saw
- The end of the demo ('Yes.' button, 'Sent.'), a caption that it is an illustration, not a recording. Then a heading 'What this is, and what it is not' with a small table: Half / Barrier / What holds it, with 'The line, expectation, the agent reads it and follows it when it can'. Then 'Neither half is a boundary, the only barrier that holds when the agent does not cooperate.' And a bullet tagged 'setting' about an organisation's admin setting a claude.ai connector tool to ask.
- Thought
- So the honest answer is: this line is a request, not a lock. The real lock is an admin setting, in claude.ai, and I am not an admin and I mostly use ChatGPT. I appreciate that it is honest, but it leaves me without a firm answer to my question.
- Question
- Is there a real lock for ChatGPT, and can I set it myself, not through an admin?
- Confusion
- The table headed 'Half / Barrier / What holds it' has one row, 'The line', and the word 'Half' is never explained; 'Neither half is a boundary' refers to a second half I never saw.
- Did
- scroll 650
- Why
- I want to read the rest of that 'stronger controls' bullet and see if it applies to me.
/policies/send-as-you/, scrolled 2600 of 4227 pixels, 390x844- Saw
- The rest of the bullet: set by the organisation's admin for each connector tool, also reaches Claude Code sessions. Then the source line, the 'tell us what happened' block, and the reading account again.
- Thought
- That is it. I have a line to paste and a test prompt. The stronger option is for admins, and for claude.ai, not ChatGPT. I will paste the line in tonight and test it. I am out of time.
- Question
- How do I know the line still works after ChatGPT updates?
- Did
- end
- Why
- I have what I came for: a line to paste into ChatGPT custom instructions and a way to test it, with the honest caveat. Four minutes are nearly up.
The interview
- In one sentence, what is this site?
- A shop of short pasteable rules and small settings that tell an AI assistant what it must not do without asking, where you pay a few pence to read each page.
- Was any of it for you? Which part?
- Yes, one page: 'Nothing is sent in your name until you have seen it'. It is my question almost word for word, for ChatGPT custom instructions. The home page and the first card on the list were not for me.
- Did you believe it? What made you believe it, or not?
- Mostly. The before-and-after test prompt and the mock-up made it feel real, and it was honest that the line is only an expectation and not a lock. That honesty made me trust it more, but it also left me unsure it will really hold with the Gmail connector.
- Where did you have to guess?
- What the 50 pounds was and why it dropped as I scrolled. What a 'hook' is. Whether 'follows it when it can' means it sometimes sends anyway. And what 'Half' meant in the table.
- What did you still not know at the end?
- Whether ChatGPT with the Gmail connector really obeys the line, whether there is a proper lock for ChatGPT that I can set myself, and whether the 50 pounds is real money I could be charged.
- Was the credit and the price of each page right for what you got?
- 50p for the page I needed felt fair, and it said so at the top. But the home page and first list card did not say the price of reading, and I paid in small amounts as I scrolled without being told on that screen. I would have paid the 50p for a firmer answer.
- Would you install a behaviour policy from here, today? If not, what would have to change?
- I would paste the line into custom instructions tonight, since it is just text and not an install. I would not install anything that needs a terminal or a hook. To go further I would need the page to say plainly whether it holds for ChatGPT with Gmail, and a way to check it myself.
- One change. What is it?
- Put my question first on the home page: 'Stop ChatGPT sending email as you', in plain words, instead of a headline about hooks and Claude Code.
- Would you send this to somebody else? Who?
- Yes, to the other product manager on my team who also lets ChatGPT draft her emails, but I would send her the link to the email page, not the home page.
- Rate the experience from 1 to 5, and say why in one sentence.
- 3: the email page gave me what I came for and was honest about its limits, but the first screen was not for me and it took four taps and a lot of skipped jargon to reach it.
Findings, and our check
| Costs | The finding | Our check | Leads to |
|---|---|---|---|
| loses a customer | The first screen speaks to developers, not to a ChatGPT user The headline paragraph names 'hook' and 'Claude Code' and the example is about keeping secrets. A non-developer who came with an email worry meets two unknown words in the first paragraph and nothing that says ChatGPT or email. She only continued because a button named 'Pick a different worry' offered a way out. step 1, | confirmed The first paragraph names a hook and Claude Code; ChatGPT first appears half a page down. | P1, P7 |
| costs time | The policy list leads with a Claude Code pack and a hook section before the ChatGPT ones The first screen of the list is a highlighted 1 pound Claude Code starter pack and the heading 'For Claude Code: a line and a hook'. The ChatGPT section with her exact question is only found after a scroll of about 1500 pixels. On a phone with four minutes that is a point where a visitor could leave. step 2, | confirmed On /policies/ the Claude Code pack and section come first; the ChatGPT section is about 1,500 pixels down on a phone. | P7 |
| costs trust | The balance drops while scrolling and the home page does not explain why The pill in the top bar goes from 50.00 to 49.99 on a list page and to 49.87 on the next, with no explanation on those screens. The cost is explained only on the page header ('this page costs 50p read to the end') and the reading account at the bottom. A first-time visitor sees money disappearing and has to guess that reading is metered. step 3, | confirmed As for Sam. | P9 |
| costs trust | The key limit of the line is stated, but the stronger lock is for claude.ai admins only The page says honestly that the line is an expectation and not a boundary, which builds trust. But the only stronger control named is an organisation admin setting for claude.ai connectors, with a Claude source link. A ChatGPT user who is not an admin is left without an answer to whether the email can really be stopped. step 7, | confirmed The only stronger control listed for send-as-you is an organisation administrator's setting in Claude, from riskmandate.ai's rule. We do not yet know what a ChatGPT user can set for themselves. | P7 |
| costs time | Unexplained terms and a one-row table on the key page Step 1 mentions 'Claude Code, CLAUDE.md' and 'expectation'. The 'What this is, and what it is not' table has a column headed 'Half' with a single row, and the text then says 'Neither half is a boundary', which refers to a half that is not on the page for a ChatGPT reader. It is a small stumble on the page that otherwise works. step 7, | confirmed Our generator wrote 'Neither half is a boundary' on pages with only one half. | P13 |
The whole study: study S1. Every review and proposal: the reviews.