"Refactor this function to use async/await instead of promises."
That line is what Glaido pastes into a coding tool after a developer says, out loud, "refactor this function to use uh, async await instead of promises". The hesitation is gone, and the term is written the way the code writes it. It is a small example of a larger problem: what people say and what a model can follow are rarely the same thing.
Glaido is a dictation app for macOS, Windows 11 and Linux. The user holds one key, talks, and the cleaned text is pasted into whichever field has the cursor, including Cursor, Claude Code, Slack and email. The company's website is glaido.com.
Why a plain transcript is a bad prompt #
People restart sentences, hesitate and leave names half-said, and a plain transcript keeps all of it. The async/await line above, left untouched, still contains "uh" and writes "async await" as two words.
Pasted into Claude Code, it asks the model to spend its first step repairing the sentence before it can act on it.
SAP keeps promoting its Joule AI tools but still hasn't published numbers to test them
Bloomberg says SAP has not published a baseline, method or period for its claim of roughly 20% productivity gains among 110,000 staff using Joule. The company's €2 billion AI efficiency target runs to 2028, while customers raised accuracy and price complaints in February. - SAP Joule AI performance metrics - why SAP withholds Joule AI results
Identifiers are where it breaks down most. Kubernetes, async/await and PostgreSQL each have one correct spelling, and speech carries neither the slash nor the capitals. If those are wrong in the prompt, the model has to guess, and in the middle of a refactor a guess is how a small edit turns into a wrong one.
From a rambling remark to an instruction #
The short case is easy to follow. A working prompt is longer, and its spoken version is messier. Glaido's own example has a developer saying, with the cursor in Cursor or Claude Code: "okay so the webhook handler, uh, it should verify the stripe signature before it parses the body, and if the signature fails just return a 400, don't log the payload".
What gets pasted is: "The webhook handler should verify the Stripe signature before it parses the body. If the signature fails, return a 400. Do not log the payload."
The speaker described the change, and the paste is the instruction that goes to the model.
The company is careful about where this stops. Its own FAQ draws the line between agentic coding, where users dictate prompts and instructions, and manual coding, where they use the keyboard.
Voice is a poor way to produce a diff. It is a useful way to produce the sentence that asks for one.
The same words, pasted differently per app #
The cleanup is not one fixed style. Glaido lets each app carry its own formatting rules, and the company's example aims the same webhook change at a colleague instead of a model. Spoken once, it is: "hey can you look at the webhook thing, the signature check has to happen before we parse, otherwise we might log junk if it's fake".
Airbnb says its own AI travel agent will book trips on and off its platform Blecharczyk told Bloomberg's The Pulse the agent will eventually book beyond Airbnb's own listings. The company's third-quarter letter targets a first-half 2027 rollout, and AI already resolves about 40% of its support issues without a human. - Airbnb AI travel agent booking trips - Airbnb AI agent off-platform bookings
Slack does not need a formal paragraph. With that app set to the casual style, which keeps capitals and uses fewer full stops, the remark can come out closer to: "Can you look at the webhook, the signature check has to happen before we parse or we might log junk if it's fake."
Email is different. According to Glaido, it has a built-in rule of its own, started on Gmail and Outlook on the web.
A short custom prompt asking for complete sentences can turn the same remark into:
"Could you review the webhook handler? The signature check needs to happen before the body is parsed. Otherwise a failed request may end up in the logs."
The rules are set once. An app with no rule of its own follows the default style, Slack follows the rule attached to it, and the hotkey is the same one used in the editor.
A dictionary for names and technical terms #
Names are the other half of the cleanup. Glaido's dictionary lets users add names and technical terms once, and the app spells them that way from then on, so adding Kubernetes means that is the spelling that appears. A correction can also swap a spoken shortcut for the full name.
The example Glaido uses is saying "Postgres" and getting "PostgreSQL". It is one entry that applies in every app.
What has to be true for voice to beat typing #
The case for dictating prompts rests on three conditions, none of which belongs to a single editor plugin.
First, the same key has to work in every text field, so Cursor, Claude Code, Slack and Gmail share one habit. Second, a dictionary has to hold the names and technical terms a person actually uses, so the cleanup does not correct them into something else. Third, the tone has to be allowed to change with the app, because a prompt for a coding agent, a Slack message and an email are not the same kind of sentence.
Glaido is built around that combination: one shortcut, a dictionary and formatting that can be set per app. The install also includes a few small desktop utilities, among them a colour picker, clipboard history and text snippet expansion. They have nothing to do with getting a prompt onto the screen.
The speed claims, and whose they are #
The company's website puts the time saved at 20 or more hours a month. It says that two to four hours of typing a day falls to under an hour, and that most people reach three to five times their typing speed within a few days.
It also calls agentic prompts about five times faster to dictate, and its dashboard example shows one user at 159 words per minute, labelled 3.1 times typing speed.
These are Glaido's own figures, and the last of them is a single example from its site rather than an average. They are worth reading as the company's claims about its product, not as independent measurements.
History on the device, speech still processed #
Privacy is the point where a tool like this gets its hardest questions, and Glaido's wording is specific. Its FAQ says dictation stays on the user's device as a temporary history and is not saved to the company's database.
That sentence is about the history. Producing the text still means the speech is processed, so it is best read as a local history, and anyone weighing the tool for a security review should look up the details first.
Glaido also states that it is ISO/IEC 27001:2022 certified and GDPR compliant, and says the certificate, its sub-processors and a route to request the audit report are published in its Trust Center. The company describes the product as private by design, with zero data retention.
Languages and pricing #
Glaido says it supports 100 languages and dialects in any application, with custom dictionaries and prompts on top. That matters for teams whose members dictate in different languages but share the same tools.
On price, the Basic plan is free, with 2,000 words a week in every app. Pro costs $17 per user per month, billed annually, and includes unlimited words, prioritised processing, early access to new features and prioritised support. Enterprise is priced on request and adds SSO and SAML, API access and dedicated data processing.
A simple test #
Glaido suggests a practical way to judge the idea. Pick a change you were about to type into Cursor or Claude Code, say it the way you would say it to a colleague, hesitation included, and read what gets pasted. If it is text you would send, the tool is doing its job.
The product does not ask people to speak like a prompt. It tries to turn the way people actually talk into text that a model, or a colleague, can use. More details are at glaido.com.
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.