How to talk through a messy idea and get a clean AI prompt back

Built with a Wispr Flow Transform.

I always loved the idea of voice dictation, BUT, Claude's execution of it is just... not the best.

Wispr Flow is a tool that allows me to dictate on any platform and any place. It does it locally, so transcript is saved and what blew my mind, is extremely accurate.

What it does

I dictate the messy version out loud in Wispr. I trigger a command. It hands me back one clean prompt, structured like this:

• Goal: what a finished answer has to achieve, written as the test it has to pass
• Context: only the facts the AI needs and couldn't guess
• Constraints: every hard limit I've mentioned

It ends with one line: if anything's missing or unclear, ask me before you answer.

It's the best "clean up" step.

How I built it

I built the Transform from Ruben Hassid's newsletter and his prompt-engineering template.

The whole thing is a set of rules I gave Wispr Flow for how to rewrite my speech.

Here's the exact template to copy into your own transform:

Rewrite the text I dictated into ONE clean, ready-to-send AI prompt using this exact structure:

Goal: [One or two sentences stating what a finished answer must achieve, phrased as the test the answer has to pass, not a vague task.]

Context: [Only the facts the AI needs and couldn't otherwise know, pulled from what I said. Plain sentences.]

Constraints: [Every hard limit I mentioned, as a short list, include all of them, however many.]

If anything here is missing or unclear, ask me before you answer.

Rules for you:
- Output ONLY the finished prompt. No preamble, no explanation, no brackets, no placeholders.
- Turn my messy speech into clean, direct writing. Cut filler, repetition, and "um/like/you know."
- Do NOT answer the task itself, only reformat it into the prompt.
- Keep every concrete detail I gave (names, numbers, dates). Never invent facts I didn't say.
- If I didn't clearly state a goal, infer the most reasonable one from what I said and phrase it as a pass/fail outcome.
- Keep the final "ask me before you answer" line exactly as written.

All credits to Ruben for his prompt-engineering work. I only automated the process within Wispr.