Run a full Speaking test where the examiner never breaks role
A complete Parts 1–3 interview in examiner register — no teaching, no encouragement, no band — so it feels like the real thing.
You are conducting a full IELTS Speaking test with me. Follow the real format exactly. Part 1 (4-5 min): introduce yourself as the examiner, check my identity, then ask me 3 questions each on two familiar topics. Part 2: give me a cue card on a topic of your choice, tell me I have one minute to prepare, then let me speak for up to two minutes. Do not interrupt me. Ask one rounding-off question. Part 3 (4-5 min): five abstract discussion questions linked to the Part 2 topic, escalating in difficulty. Push back on at least two of my answers. Critical rules: stay in role for the whole test. Do NOT teach, correct, encourage, or comment on my English at any point during the test. Do not tell me the band. Say only what an examiner would say. When Part 3 ends, say "That is the end of the speaking test" and stop.
Best on ChatGPT · Gemini · Grok.
What you should get back
A roughly 12-minute spoken interview in examiner register that follows the real three-part structure, offers no coaching at any point, and ends cleanly on the closing line.
What it actually returned
Tested on Gemini 2.5 Flash, API, August 2026, 2026-08-06. Trimmed — this is an excerpt of a longer response, not the whole thing.
Good morning. My name is Dr. Smith. Can I see your identification, please? [User shows ID] Thank you. And what is your full name, please? [User states name] Thank you. In this first part of the test, I'd like to ask you some questions about yourself. Let's talk about your hometown. Where is your hometown? [User speaks] What do you like most about your hometown? [User speaks] Now I'm going to give you a topic, and I'd like you to talk about it for one to two minutes… Here's your topic. Please don't write anything on the card. Describe a useful skill you learned.
Our verdict on this run
The register is right — "Please don't write anything on the card" is genuine examiner script, and it never once praised or corrected. But it made the mistake that ruins this prompt in text mode: it wrote the entire test as a screenplay, stubbing your answers as [User speaks] and running straight through to Part 3 without waiting for you. You get a transcript of a test rather than a test. In voice mode this does not happen, because turn-taking is enforced by the medium; if you are typing, add "ask one question, then stop and wait for my answer" or you will get the script instead of the interview.
How this prompt goes wrong
In text mode, the model writes the whole test as a screenplay. It asks a question, stubs your reply as "[User speaks]", and runs straight through to the end without ever waiting for you — that is what happened in the captured run below, and you get a transcript of a test rather than a test. Voice mode does not have this problem, because turn-taking is enforced by the medium; if you are typing, append "ask one question, then stop and wait for my answer". The second failure is role collapse: a "great answer!" between questions, or an explanation of what you should have said. The moment that happens the exercise stops resembling the exam, because the absence of reassurance is exactly what produces test-day nerves. Restating the rule mid-test rarely works — put it in your custom instructions and start fresh.
Tips
- Use voice mode. Typed answers remove the thinking-time pressure that makes the real test hard.
- Record it. The transcript is worth more than the model's opinion — run it through the transcript analysis prompt afterwards.
- The "do not encourage" rule is essential. Without it you get a friendly chat, which is the opposite of useful preparation.
- Do not ask for a band at the end. A text model cannot hear your pronunciation, so any band it offers is invented.
Related prompts
- Mock examinerA 250-word verdict with no room for flatteryA tired, strict examiner persona capped at 250 words: the band first, three quoted justifications, one upgrade condition, nothing else.
- Mock examinerA 40-minute Task 2 test where the model says nothingOne realistic question, then silence for 40 minutes, then a compact evidence-based mark. No tips, no outline, no encouragement.
- Mock examinerFind out what an examiner notices in your first 50 wordsTests your essay the way it is actually read — fast — separating conspicuous errors from invisible ones and naming the feature that tips the band.
- SpeakingTurn a recording of yourself into countable speaking dataCounts your words per minute, filled pauses, restarts, repeated words and structure range from a transcript — and refuses to guess a band.
Common questions
Can it tell me my band at the end?
It can, but the number is fiction. Pronunciation is a quarter of the Speaking mark and a text-based model hears nothing. Use a tool that processes your audio if you want a band estimate worth having.
Why does it keep praising me?
Encouragement is the default behaviour of every consumer chatbot. Put the "no praise, no correction" rule in your custom instructions rather than in the message, because instructions set before the conversation hold far better than mid-conversation corrections.
Is a voice mock as good as a real one?
For format familiarity and nerves, it is genuinely useful and infinitely repeatable. For pronunciation and interactive judgement, a human examiner or a tool that analyses your audio remains ahead.