Prep for your OpenAI interview inside ChatGPT. It plays the interviewer.

Photo: Clint Patterson on Unsplash

Add one URL to ChatGPT, Claude or Grok and it walks you through OpenAI's loop and gives you questions one at a time, attempt first. It reaches 2,169 questions at 60 companies.

• Add one URL and ChatGPT, Claude or Grok plays your interviewer, with the loops and questions on Land the Lab. You sign in once. There is no key to copy.
• It reaches 2,169 questions at 60 companies. 136 are reported at OpenAI and 64 at Anthropic.
• 1,058 prompts open in the chat for any member. The other 1,111, the recently asked and most asked ones, show a teaser until you unlock them.
• It gives you one question at a time and asks for your attempt first. The worked answer comes after, and only if you have unlocked it.
• 6 tools, 5 of them read only. The one that changes anything spends a credit, and only when you say yes.

Three steps in each assistant, then you prep where you already chat

Land the Lab runs a connector: a server your AI assistant can call while you talk to it (an MCP server, in the assistants’ terms). You add its URL once, sign in to Land the Lab in your browser and choose Allow. The URL is https://landthelab.com/mcp. These are the steps on our AI assistant page:

Claude

1. Open Settings → Connectors and choose Add custom connector.
2. Name it Land the Lab and paste the URL.
3. Choose Connect, sign in to Land the Lab and choose Allow.

In Claude Code: claude mcp add --transport http land-the-lab https://landthelab.com/mcp

ChatGPT

1. Open Settings → Apps → Advanced settings and turn on Developer mode.
2. Choose Create app, name it Land the Lab, paste the URL and pick OAuth.
3. Sign in to Land the Lab and choose Allow. In a chat, pick it from the + menu.

Developer mode is on paid ChatGPT plans.

Grok

1. Go to grok.com/connectors and choose New Connector → Custom.
2. Paste the URL and leave authentication on OAuth.
3. Sign in to Land the Lab and choose Allow.

Meta Muse (custom connector)

1. Open Muse's connector settings and add a custom connector.
2. Paste the URL as the remote MCP server.
3. Sign in to Land the Lab and choose Allow.

Meta Muse takes it as a custom connector today. You can remove the connection any time in your assistant or under Settings → API tokens on Land the Lab.

It reaches 2,169 questions at 60 companies, 1,524 of them coding

The assistant sees what you see on the site. That is every company page, 47 of them with a known loop, and every question, newest first. It can filter by company, type and round. 487 questions have an official worked answer.

For the labs: OpenAI has 136 reported questions, Anthropic 64 and SpaceXAI 47. Ask for a company and the assistant starts from its loop, round by round, with the questions each round asks most.

1,058 prompts are open in the chat. The 1,111 asked recently or most need an unlock

A question is locked when a company asked it in the last 6 months or it is among that company’s 10 most asked. The connector follows the same rule as the site, because it reads the same pages. On a locked question your assistant gets the title, the companies, the round, when it was last asked and a short teaser. The full prompt never leaves our server until you unlock it. At the labs, most of them are locked:

112 of OpenAI’s 136 questions (82%) are locked: most of them were asked in the last 6 months. New accounts get 3 credits. One credit unlocks one question for good, its prompt and its answers. A paid plan opens all of them. On an open question the prompt is free and the worked answer takes the same one credit.

An example: one OpenAI question, attempt first, answer after

This is an illustration of how a session goes, written from the connector’s tools and the messages it returns. Your assistant’s words will differ. The question is real: Text editor with undo and redo, reported at OpenAI and open to every member.

You

I have an OpenAI onsite in three weeks. Walk me through the loop.

Assistant · calls get_company_loop

OpenAI’s loop has 7 rounds:

1. Recruiter screen, 30 minutes
2. Technical screen: coding, 60 minutes
3. Technical screen: system design, 60 minutes
4. Onsite coding, 60 minutes
5. Onsite system design, 60 minutes
6. Project presentation, 45 minutes
7. Behavioral

Each round lists the questions it asks most. Some are locked, so for those I only see a short teaser.

You

Give me a coding question I can open now.

Assistant · calls search_questions, then get_question

Try “Text editor with undo and redo”, last asked Mar 2026. [The full prompt.] Before you write code, tell me your approach.

You

[Your approach, then your code.]

Assistant

Pushes back as an interviewer would: asks about edge cases and the cost of each operation, and gives a hint before any solution.

You

Show me the official answer.

Assistant · calls get_answer

The answer is locked. Unlock this question’s answers with 1 credit (you have 3). Want me to spend one?

You

Yes.

Assistant · calls unlock_question, then get_answer

Unlocked “Text editor with undo and redo” for good: prompt and answers. 2 credits left. Here is the official answer, and where your attempt differs from it.

The assistant is told to work this way: ask the company and date, walk the loop, give one question at a time, let you try first, and only then fetch the answer. Those are instructions to the model. The locks are enforced on our server.

What it won’t do, and the limits

• It can’t read a locked prompt or answer. It is told never to guess or rebuild one. It offers an unlock, links the question page or moves to an open question.
• It changes nothing but your credits, and only when you ask. 5 of its 6 tools only read. The other is unlock_question.
• Work in the chat isn’t saved to Land the Lab. Your canvas, the code runner with tests by level and the AI tutor are on each question’s page. Diagrams in an answer show as a link to that page.
• It reads your prep, it doesn’t set it. It sees your target company, interview date, progress and credits. You set the target on My prep.
• Too many questions too fast and it has to wait, the same limit as opening question pages on the site.
• The replies are your assistant’s. Land the Lab sends the loops, questions and answers. Your assistant’s model writes what you read, under your assistant plan’s own limits.

What to ask first

1. “I have an Anthropic onsite in three weeks. Walk me through the loop.” Lays out the loop round by round, with the questions each round asks most.
2. “Give me OpenAI's most asked coding question. Let me try it before you show the answer.” Plays the interviewer: one question at a time, your attempt first, then the worked answer if you've unlocked it.
3. “What should I work on next in my prep?” Reads your prep target on Land the Lab and picks the next question you haven't done.

Developers and reviewers can read each tool’s inputs and outputs in the tool documentation.

Methodology

Counts from Land the Lab’s data on Oct 21, 2026: 2,169 questions (merged duplicates left out), 60 companies with at least one, 47 with a known loop. A question counts as locked when it was premium that day (6 months, top 10 per company); a member who unlocked it, or has a paid plan, sees its prompt. The tools, their messages and the setup steps are the ones the connector and the AI assistant page use as built.

Limits: the example conversation is an illustration, not a transcript. Assistants change their connector settings and plans; follow the steps on the AI assistant page, which we keep current.

How we made this: the counts are generated from our data; the text was drafted with AI and checked against the product.

Originally published at landthelab.com, where every company's interview loop is mapped round by round.