Prompts that measure what a transcript can actually show.
Speaking is the skill AI helps with least honestly. A text model cannot hear you, and pronunciation is one of the four Speaking criteria — a full quarter of your mark. Any chatbot that reads a transcript and hands back a Speaking band has invented 25% of it. The prompts here are written around that limit rather than pretending it away.
What a transcript genuinely supports is measurement: words per minute, filled pauses, self-corrections, restarts, your ten most repeated content words, the grammatical structures you used and — more usefully — the ones you never used once. Those are countable, trackable week to week, and they map onto Fluency and Coherence, Lexical Resource and Grammatical Range directly. The transcript analysis prompt below is deliberately forbidden from estimating a band precisely because that is the part it cannot know.
The second thing AI does well here is pressure. Part 3 is where Band 6.5 becomes Band 7.5, and the deciding skill is extending and defending a position rather than deploying vocabulary. A model instructed to challenge every general claim with a counter-example, press you when you hedge, and never once encourage you produces something close to the real experience — but only if it stays in role, which is the failure mode documented on each prompt page.
The third thing worth knowing is what these prompts cannot fix. A model reading your transcript sees vocabulary and grammar, so it will tell you to use more precise words and more varied structures — advice that is true of almost everyone and therefore not very useful. It cannot hear that you trail off at the end of sentences, that your intonation flattens when you are nervous, or that a particular consonant cluster is costing you intelligibility. Those are the things that separate a 6.5 from a 7.0 for a great many candidates, and no amount of prompt engineering will surface them from text.
A workable routine: run a cue card in voice mode with the examiner prompt, record it, transcribe it, then run the transcript through the analysis prompt. The counts go in a note — words per minute, filled pauses, restarts, and the structures you did not use. Four weeks later, do the same card again and compare the two sets of numbers. That loop is worth more than any single band estimate, because it measures change rather than level, and change is the only thing you actually control between now and test day.
Counts your words per minute, filled pauses, restarts, repeated words and structure range from a transcript — and refuses to guess a band.
Best on ChatGPT · Claude · Bands 6.0–7.5
Produces three spoken-register answers to one cue card at ascending bands, then names the concrete moves that separate each level from the next.
Best on Claude · ChatGPT · Bands 6.0–8.0
Six escalating Part 3 questions with real counter-examples and follow-ups, then a report on where your answers were too short or unsupported.
Best on ChatGPT · Gemini · Bands 6.5–8.0
It can assess what is in a transcript — fluency markers, vocabulary range, grammatical range, coherence — but not pronunciation, which is a quarter of the mark. A band from a text model is missing that quarter entirely. Tools that process your actual audio are a different matter.
For format familiarity, question exposure and Part 3 pushback, genuinely yes, and it is infinitely repeatable at no cost. For pronunciation feedback and the interactive judgement of a real examiner, it is not a substitute.
Your phone's voice typing, a browser dictation feature, or the transcription in most note-taking apps all work. Do not clean up the result — the "um"s, restarts and self-corrections are exactly the data the analysis prompt counts.