Most people type worse prompts than they’d say out loud. When you type, you compress. You leave out the context because it’s tedious to write, you skip the constraints because you’ll “fix it in the follow-up”, and you end up with a one-line question and a three-turn conversation to get what you wanted.
When you talk, you ramble, and rambling is full of exactly the detail a model needs. The trick is turning the ramble into a prompt without losing that detail.
What Claude actually wants
Strip away the prompt-engineering folklore and a good prompt has four parts.
The ask. One line. What you want back.
Context. What the thing is for, who it’s for, what already exists.
Constraints. The “no”, the “must” and the “keep it”. These are the bits you’d naturally say while talking (“and don’t use Tailwind”) and naturally forget while typing.
Output. The shape you want: a file, a list, three options, a paragraph.
Say these four things and Claude’s first answer is usually the right one. Miss them and you spend the next ten minutes correcting.
The problem with just dictating
Ordinary dictation gives you a transcript. Cleaned up, maybe, but still in the order you said it, with the ask buried in the middle and the constraints scattered through. Claude copes, but it’s working harder than it should, and so are you.
Here’s a real ramble:
ok so i need a landing page for the app, it’s a mac dictation thing, dark, keep it short, no stock photos, and uh don’t use tailwind, just one html file
And here’s the same thing as a prompt:
Build a landing page for a macOS dictation app.
Context: hold a key, talk, and it types clean text into any app. The audience is Mac users who write all day.
Requirements: dark theme, short copy, no stock photography, no Tailwind or frameworks, one self-contained HTML file with inline CSS.
Return the full file, then a two-line summary of the design choices.
Same information. The second one gets a usable file on the first try.
Doing it without typing
You can do this restructuring by hand, but then you’re back to typing. The point of Speaky Box is that the Smart key does it for you: hold the key with Claude in front, say the ramble, let go, and the structured version is typed into the message box. In Mail the same key writes an email, in Slack a message. It reads which app you’re in.
A few things that make it work better:
Say the constraints out loud, even the obvious ones. “Keep it under 200 words.” “Don’t change the function names.” They cost you a second to say and they’re the difference between one answer and three.
Mention the audience. “This is for a customer who’s already annoyed” changes the whole answer.
Name the output. “Give me a table”, “give me the file”, “give me three options and pick one.” Claude will do what you ask, so ask.
Add your terms to the vocabulary. Product names, people, libraries. Once they’re in Speaky Box’s Words list they’re spelt right every time, in every app.
When to use plain dictation instead
Not everything you type into Claude is a prompt. Replies inside a conversation (“yes, the second one, but shorter”) don’t need structure. That’s what the Dictate key is for: your words, tidied, nothing rearranged. Two keys, two jobs, and you stop thinking about it after a day.