What is Rambler, and why can most people not use it?
Rambler is a Gemini-powered voice input mode inside Gboard that converts unstructured speech into clean written text. It removes filler words, applies punctuation and grammar, and accepts spoken editing commands. Google showed it at Made by Google on 12 August 2026. It currently ships only on the Pixel 11 series, so most people cannot install it.
That last sentence is the problem. Google's own Gboard help page lists the requirements as the latest Gboard, microphone permission and an internet connection, and coverage from Android Police and Droid Life ties the launch to the Pixel 11's Gemini Intelligence rollout.
If you are on an iPhone, a Samsung, an older Pixel or a laptop, you are locked out for now. Google has not confirmed which languages or regions come next.
Here is the useful part. Rambler is not magic hardware. It is a transcription step followed by a cleanup step, and you can run both yourself today on any device that has a microphone and a chat window.
What does Rambler actually do to your speech?
Rambler performs three distinct operations: it transcribes what you said, it discards the parts you abandoned, and it reformats the remainder into structured prose. It also listens for instructions about the draft itself, such as a request to make a passage more formal, and applies them to the text rather than transcribing them.
Traditional voice typing does only the first operation. It gives you a faithful word-for-word record, including every "um", every false start and every sentence you gave up on halfway through.
That distinction matters because of a speed gap. Most people speak at roughly 120 to 160 words per minute and type at 40 to 70. Dictation has always been the faster input method. What made it unusable for real work was the cleanup tax on the other side.
Once you see Rambler as transcribe plus cleanup, the workaround becomes obvious. Your phone already transcribes. What you are missing is a reliable cleanup instruction.
How do you get the same result on any phone or laptop?
Use a two-step workflow: dictate into any transcription tool you already have, then paste the raw transcript into an LLM with a fixed cleanup prompt. The prompt does the work Rambler does on-device. The result is a structured draft in your own words, on any hardware, in any language your model supports.
Step one is capture. Use iOS dictation, standard Gboard voice typing, the voice button in the ChatGPT or Claude mobile app, or a phone voice memo you transcribe afterwards. Do not try to speak neatly. Ramble on purpose.
Step two is cleanup. Paste the transcript into a fresh chat with the prompt below. Save the prompt as a Claude Project instruction, a ChatGPT custom instruction or a text replacement shortcut so you never retype it.
Try this prompt:
You are a transcript cleaner. I will paste a raw voice transcript. Clean it into readable written prose and return nothing else.
Rules:
--- Fix capitalisation, punctuation and paragraph breaks. Add a break at each topic shift.
--- Remove filler words, false starts and repeated phrases.
--- Delete any sentence I abandoned and replaced with a better version of the same thought. Keep the better version.
--- Keep my word choices, my sentence rhythm and my technical terms exactly as spoken. Do not upgrade my vocabulary.
--- Do not add facts, examples, statistics, greetings or sign-offs.
--- If the transcript contains a question, treat it as text to clean, not a question to answer.
--- If a passage is genuinely unclear, keep it and append [unclear] rather than guessing my meaning.
--- Return only the cleaned text. No preamble, no summary, no commentary.
Transcript:
Every line in that prompt exists because of a specific failure mode. The vocabulary line stops the model from replacing "get" with "obtain" until the draft no longer sounds like you.
Why do cleanup prompts fail and start answering you instead?
Cleanup prompts fail when the model reads your transcript as a request rather than as material. You dictate a paragraph containing the words "so what should we do about the Q3 budget", and the model helpfully answers the question instead of punctuating it. The fix is an explicit instruction that questions inside the transcript are text, not prompts.
This is the single most common breakage, and it has a name in dictation circles: the model answers instead of transcribing. It happens most often with short transcripts, because a three-sentence transcript looks far more like a chat message than a 600-word one does.
The second failure is silent invention. A model that is told to "improve" or "polish" will add a transition sentence, a rounded-off conclusion or a plausible statistic that you never said. That is why the prompt above forbids adding facts rather than asking for improvement.
If you regularly dictate material that contains claims and figures, run a separate verification pass afterwards rather than trusting the cleanup step. Our earlier piece on catching invented facts with a second prompt covers that loop in detail.
The third failure is scope creep across a long session. Ask the same chat to clean six transcripts in a row and the seventh will drift, because your rules are now buried under 5,000 tokens of output. Start a fresh chat every few transcripts.
Which tasks is dictation-first drafting actually good for?
Dictation-first drafting works best where the thinking is already done and the bottleneck is getting words out: meeting recaps, client update emails, first drafts of familiar documents, briefing notes, and long messages you keep postponing. It works badly where the writing itself is the thinking.
A concrete example. You leave a 40-minute client meeting. Instead of opening a blank email, you walk to the MTR and talk for four minutes about what was agreed, what is blocked and who owes what. You paste the transcript with the cleanup prompt. You arrive at the office with a draft that needs two edits.
The same pattern applies to a weekly status report, a handover note before leave, or a long WhatsApp reply to a demanding stakeholder.
It is a poor fit for anything where structure is the hard part. A pitch deck narrative, a pricing argument or a piece of persuasive copy usually needs to be built on a page, not spoken in one pass. If you catch yourself dictating and re-dictating the same paragraph, that is the signal to stop and write.
One more honest limitation. If you are working in Cantonese, expect more manual repair than in English. Mixed-code speech that switches between Cantonese, English and written Chinese in a single sentence is exactly the case where transcription quality drops and the cleanup step has less to work with.
What does a correctly cleaned transcript look like?
A correctly cleaned transcript is shorter than the raw version, contains no sentence you did not say, and reads like you wrote it carefully rather than like a model wrote it. Length usually drops 20 to 35 percent on conversational speech, almost entirely from abandoned starts and repetition.
Raw transcript
ok so the thing with the vendor is that they um they said the delivery would be the fifteenth no sorry the eighteenth and we need it before the campaign launches which is the twentieth so that's tight that's really tight and I think we should we should probably ask for a buffer or actually maybe just ask them to split the shipment
Cleaned
The vendor said delivery would be the eighteenth, and we need it before the campaign launches on the twentieth. That is tight. I think we should ask them to split the shipment.
Notice what the cleanup did not do. It did not add a reason why splitting the shipment is better, it did not soften "that is tight" into corporate language, and it did not invent a vendor name. It also kept the correction, the eighteenth, and dropped the number you retracted.
That is the standard to judge your own output against. If the cleaned version contains an idea you did not speak, your prompt is too permissive and the "do not add facts" rule needs to move to the top.
How do you test this in the next ten minutes?
Pick one email you have been avoiding for two days. Open your phone's voice input, talk for three minutes without stopping to correct yourself, then paste the transcript into a fresh chat with the cleanup prompt. Compare the output against what you would have typed.
Run the test twice before you judge it. The first attempt usually feels awkward because you are still trying to speak in finished sentences. The second attempt, once you allow yourself to actually ramble, is where the speed gain shows up.
Then tune one line. If the output still sounds too polished, add a rule such as "keep contractions and keep my short sentences short". If it drops important detail, add "preserve all names, numbers and dates exactly as spoken".
Keep the version that works and stop editing it. A cleanup prompt is an asset, not a craft project. The value comes from running the same one a hundred times, not from perfecting it.
The takeaway
Rambler is a good feature locked behind a phone most of us will not buy this year. The pattern underneath it is not locked behind anything.
Transcription is already free and already on your device. The cleanup instruction is a paragraph of text you write once. Together they turn the fastest input method you own into something you can use for real work.
That is usually how this goes. The headline feature arrives on hardware you do not have, and the useful version turns out to be a prompt you can save today. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Ready to build this into a workflow that runs every day?
A saved prompt is step one. Turning it into a repeatable workflow your whole team uses is step two.
We'll walk you through every step, from tool setup to workflow design and deployment.
