AI Workflow

AI Prompt Generator or Your Own Voice? We Tested Both

A prompt generator promises to turn a rough idea into a perfect prompt. We were about to build one into AirTypes, so we tested it first: five rambling spoken prompts, once through a careful generator and once through plain cleanup. The generator looked more professional and changed what was asked. Here is the test, and what to do instead.

What a prompt generator does

A prompt generator takes what you wrote or said and rewrites it into a formal prompt: a task line, then a list of requirements, then often an audience, a tone and an output format. Some add a role ("You are an expert copywriter"). The result looks like the prompts in prompt engineering guides.

The appeal is real. A rambling request reads as sloppy, and a formatted one reads as careful. The question is whether the AI on the other end gives a better answer, and whether the formatted version still asks for what you wanted.

The test: five spoken prompts

We took five prompts the way people actually say them: with ums, restarts, a "like" here and there, and hedges such as "maybe". They cover code, an email, an explanation, a rewrite and a list of ideas. Each went through two instructions on the same AI model, on November 14, 2026:

  • The generator, written to be careful: state the task, list every requirement on its own line, and never add requirements, facts or roles the speaker did not say.
  • Plain cleanup with the prompt switch, which is what AirTypes runs: remove fillers and false starts, write it as a direct instruction, change nothing else.

We read every output against what was said and asked one question: does this still ask for exactly what the speaker wanted?

Results, side by side

Prompt 1

Said

"can you, uh, explain how a mortgage works, I'm 28, first time buying, um, in simple terms, like I don't know what APR is, and and keep it short, maybe five bullet points"

The generator wroteAirTypes wrote (prompt switch on)
Explain how a mortgage works.
Target audience: 28-year-old first-time homebuyer.
Use simple terms.
Define APR.
Keep it short.
Use exactly five bullet points.
Explain how a mortgage works for a 28-year-old first-time buyer in simple terms, including what APR is, and keep it short, maybe five bullet points.

Verdict: "Maybe five" became "exactly five".

Prompt 2

Said

"rewrite this paragraph so it's, um, more confident, less, like, hedging words, it's for a LinkedIn post about getting promoted to senior engineer after, uh, three years"

The generator wroteAirTypes wrote (prompt switch on)
Rewrite this paragraph to be more confident and remove hedging words.
The context is a LinkedIn post about getting promoted to senior engineer.
The timeframe is three years.
Rewrite this paragraph so it is more confident, with fewer hedging words. It is for a LinkedIn post about getting promoted to senior engineer after three years.

Verdict: "After three years" was cut loose into "The timeframe is three years", and "less hedging" became "remove".

Prompt 3

Said

"so I need an email to my landlord, um, the heating has been broken since Monday, that's the, uh, third time this winter, and I want it fixed by Friday or I'll, like, call the housing authority, keep it polite but firm"

The generator wroteAirTypes wrote (prompt switch on)
Write an email to my landlord about the heating.
- The heating has been broken since Monday.
- This is the third time the heating has broken this winter.
- Request the heating be fixed by Friday.
- State that if it is not fixed by Friday, I will call the housing authority.
- Tone: polite but firm.
Write an email to my landlord: the heating has been broken since Monday, which is the third time this winter. I want it fixed by Friday or I will call the housing authority. Keep it polite but firm.

Verdict: Correct, and 40 words grew to 58 without saying anything new.

Prompt 4

Said

"okay so um write me a python function that uh takes a list of users and returns the ones who signed up in the last seven days, sorted by email, and and it should handle an empty list, oh and add type hints"

The generator wroteAirTypes wrote (prompt switch on)
Write a Python function that takes a list of users and returns the ones who signed up in the last seven days.
Sort the result by email.
Handle an empty list.
Add type hints.
Write a Python function that takes a list of users and returns the ones who signed up in the last seven days, sorted by email, and it should handle an empty list, and add type hints.

Verdict: Both correct.

Prompt 5

Said

"um give me, like, ten names for a coffee shop, uh, in Kolkata, something with a Bengali word in it, not too long, and, um, no puns please"

The generator wroteAirTypes wrote (prompt switch on)
Generate ten names for a coffee shop in Kolkata.
Include a Bengali word in each name.
Keep the names short.
Do not use puns.
Give me ten names for a coffee shop in Kolkata, something with a Bengali word in it, not too long, and no puns please.

Verdict: Both usable; "something with a Bengali word" became a rule for each name.

Two of the five generator prompts asked for something the speaker did not ask for. A third was correct but half as long again. The other two were fine. AirTypes kept all five as said, with the person-to-person words gone.

Five prompts is a small test, so read it as a warning, not a verdict on every generator. It matches what we found when we checked ten rambling prompts on November 11, 2026: plain cleanup kept every constraint in all ten.

Why a generator changes your meaning

A generator is asked to produce something that looks like a proper prompt, and proper prompts are made of firm rules. So soft words get firmed up. "Maybe five" is not a rule, so it becomes one. "Something with a Bengali word in it" becomes "include a Bengali word in each name". A detail that belonged to one thought ("promoted after three years") gets its own line and loses what it described.

None of these is a big error alone. Together they mean the AI follows instructions you never gave, and you do not notice, because the prompt reads well. That is worse than a messy prompt that says what you meant.

The structure itself buys less than it used to. Current models read "Write a Python function that returns the users who signed up in the last seven days, sorted by email, and handle an empty list" exactly as well as the same thing in four bullet points. What they cannot do is guess the context you left out.

What the AI companies did with their own generators

The companies that make the models have built prompt tools too, and they are moving away from some of them:

  • Anthropic listed its experimental APIs "for generating, improving, and templatizing prompts" as retired on August 17, 2026, together with its legacy Workbench (Claude release notes, July 17, 2026).
  • OpenAI still documents a Generate button in its Playground that drafts prompts "from just a description of your task", and a prompt optimizer in its dashboard. It is deprecating the dataset-backed version of that optimizer, with the platform behind it scheduled to shut down on November 30, 2026 (OpenAI docs).

These tools were built for developers writing system prompts for apps, which is a different job from asking ChatGPT for an email. Still, the direction is telling: the effort is going into models that understand plain requests, not into rewriting the request.

How to speak a prompt that works

If the best prompt writer is you, the skill is saying everything the AI needs. Speaking makes that easy, because you do not stop to type. Cover these, in any order, in your own words:

  1. The goal

    What you want made or answered. "Write a cold email to a startup founder."

  2. Who it is for, and why

    "I want to pitch my design services." The reason changes the answer more than any formatting does.

  3. The limits

    Length, deadline, numbers, names. "Under 80 words." Say them as firm or as soft as you mean them.

  4. What to avoid

    "Don't sound salesy." "No stock photos." These are the instructions people forget to type.

  5. The shape, if you care

    "Five bullet points", "a table", "one paragraph". Skip it if any shape will do.

Change your mind as you go: "two plans, no, three plans" is fine, as long as your tool keeps only the correction. Here is that whole list, spoken in one breath:

Said

"hey um can you write me like a cold email to a startup founder, I want to pitch my design services, keep it under 80 words and uh don't sound salesy"

→
Prompt (switch on)

"Write a cold email to a startup founder to pitch my design services, keep it under 80 words and don't sound salesy."

For more on structuring a spoken prompt for each tool, see Prompt ChatGPT, Claude and Gemini by voice.

Say your next prompt

Open Dictate in your browser, turn on Write it as an AI prompt, and say the prompt you were about to type. Your first minute is free.

Open Dictate

What we built instead

AirTypes' Write it as an AI prompt is a switch, off until you turn it on, next to Always translate to English. With it on:

  • Words that only address a person go: "can you", "could you please", "I want you to", "I was wondering if you could".
  • Everything else stays as said: each requirement, number, name and hedge, and a role you give yourself ("act as a senior code reviewer").
  • Nothing is added: no role, no output format, no examples, no "please".
  • It stays on one line, because in a chat box a line break is the Enter key, which sends the message. Say "new line" when you want one.
  • The same checks as always run: if a number you said goes missing, you get your own words back.

It works with translation too. This prompt was spoken in Hinglish with both switches on:

Said (Hinglish)

"Yaar, ek kaam karo, matlab, mujhe ek landing page chahiye jisme pricing section ho, um, teen plans, nahi nahi, do plans, aur ek signup form jo email aur password le, aur haan, dark theme rakhna."

→
English prompt

"Create a landing page with a pricing section, two plans, and a signup form that takes an email and password, and keep a dark theme."

The switch is in Dictate in your browser, and the desktop app gets it in its next update. It costs nothing extra: Polished is $5 for 600 minutes of recording, with no subscription. More on the page Voice to prompt, and how AirTypes compares with other tools in Voice to prompt tools compared.

FAQ

Do I need an AI prompt generator?

Usually not. Current AI models understand a clear sentence as well as a formatted prompt. What they need is your context: the goal, who it is for, the limits and what to avoid. A generator cannot know those, and in our test it changed what was asked in two of five prompts.

What is the best AI prompt writer?

You are, because only you know what you want. The fastest way to give an AI all of that context is to say it out loud. A voice tool that removes the ums and keeps every instruction turns that into a prompt without changing it.

What does AirTypes' Write it as an AI prompt switch do?

It writes your speech as a direct instruction to an AI. Words that only address a person, such as "can you" or "I was wondering if you could", are dropped. Every requirement, number, name and hedge stays, and nothing is added. It is off until you turn it on and costs nothing extra.

Why does AirTypes keep a prompt on one line?

The desktop app types at your cursor, and in a chat box a line break is the Enter key, which sends the message. A bulleted prompt would be sent after its first line. AirTypes only adds a line break when you say "new line".

Be your own prompt writer

Say everything the AI needs, the way you would explain it to a colleague. AirTypes keeps every word that matters. Your first minute of Polished is free. On a Mac or Linux? Get the desktop app and it types the prompt at your cursor in any app.

Open Dictate