A good spoken prompt has four parts, said in order: goal, context, constraints, output. If you can say those four sentences, dictation will give you better prompts than typing does, mainly because you will say more than you would bother to type. This post gives you a template to say aloud, shows before and after examples, and covers cleanup and privacy. Tool choice is a separate question; see the dictation tool comparison.
Why spoken prompts go wrong
When you type, you edit as you go. When you speak, you think out loud. Three habits ruin spoken prompts:
- Stacking. "Also, while you're at it" turns one request into four.
- Trailing context. The key constraint arrives in the last sentence, after you have already described the wrong goal.
- Vague referents. "Make it better" and "like the other one" mean nothing to a model that cannot see inside your head.
The fix is not to speak slowly. It is to speak in a fixed shape.
The four-part template to say aloud
This is an untested template, built from standard prompt structure (see the prompt elements lesson for the written version). Say it nearly verbatim:
"The goal is [one outcome]. Here's the context: [who it's for, what exists already, what you've tried]. Constraints: [what to avoid, length, tone, tools or files not to touch]. Give me the answer as [format]."
Four spoken sentences, thirty to sixty seconds. The order matters: goal first lets the model interpret everything after it, and the output instruction last is easy to remember.
Add two optional lines when the stakes are higher: "Before you answer, tell me what you're assuming," and "Ask me up to three questions if anything is unclear."
Before and after examples (illustrative)
These are hand-written to show the pattern. They are not captured transcripts.
Writing task, raw dictation:
"Okay so I need an email, um, to a client, basically they're late on the invoice again and I don't want to be rude but also it's been like six weeks, and, yeah, make it professional."
After applying the template:
"The goal is a reminder email about an overdue invoice. Context: the client is a long-term customer, the invoice is six weeks overdue, and this is the second reminder. Constraints: polite but firm, under 120 words, no threats, and mention that I can resend the invoice. Give me the answer as a subject line and body."
Coding task, raw dictation:
"The signup thing is broken, can you fix it, it's the thing that, I think it's the form."
After:
"The goal is to fix the signup form so it shows an error when the email is already registered. Context: right now it submits and shows a blank page. I think the problem is in the form submit handler. Constraints: don't change the API, and keep the existing validation library. Give me a short explanation of the cause first, then the change."
Notice what the second versions share: one job, concrete nouns, a stated format. For coding agents, pair this with the advice in vibe coding prompting.
Dictate meaning, type symbols
Speech is good at intent and bad at syntax. Dictate the goal, context and reasoning. Type file paths, regexes, commands, version numbers and anything where one wrong character matters. A practical hybrid: speak the paragraph, then click into it and type the two or three exact identifiers.
Say numbers and units plainly ("one hundred twenty words", "under two seconds") and read them back in the transcript. Spell unusual names once, either by letter or by adding them to your tool's custom dictionary if it has one.
Cleaning up a transcript in 30 seconds
Many dictation tools already strip filler words and add punctuation. You still need a quick pass. My checklist:
- Negations. Search for "not", "don't", "never". Dropped negatives are the costliest error.
- Names and identifiers. Library names, function names, people, products.
- Numbers. Dates, limits, counts.
- Stray requests. If you said two jobs, split them into two prompts.
- The output line. Make sure the format instruction survived.
If the transcript is long and messy, you can paste it into the model with a cleanup instruction. This is an untested template:
"Below is a dictated prompt. Rewrite it as a clear prompt with sections: Goal, Context, Constraints, Output format. Don't add information I didn't give. If something is missing, list the missing items as questions instead of guessing."
The last line matters. Without it, a model may fill gaps with plausible inventions, which is the opposite of what you want from a rewrite.
Privacy cautions for spoken prompts
Voice adds exposure that typing does not: audio of your voice, background conversation, and in some tools, text read from your screen for context. Treat it like sending the content to a third party unless you have confirmed otherwise.
- Know where audio is processed. Wispr Flow's privacy page states transcription always happens in the cloud and that audio is deleted after transcription unless cloud storage is on. Superwhisper's docs say that with local models nothing leaves your device. Read your own tool's page, dated the day you read it.
- Never speak secrets. API keys, passwords, tokens, private keys, customer personal data.
- Check training and storage settings. Some tools have a setting for using your data to improve models. Know where it is and what it is set to.
- Watch the room. Dictating client details in a cafe is a leak of a different kind.
- Work policy first. If your employer or client restricts cloud tools, local-only dictation is the safer route, and you should ask before assuming.
For a worked developer setup that follows these cautions, see the voice coding workflow with Claude Code and Cursor. If you are new to agent tools, Claude Code for non-coders is a gentle start.
A ten-minute practice routine
- Pick three prompts you typed recently that got poor results.
- Re-say each using the four-part template. Do not read from a script.
- Compare outputs with the typed originals.
- Note which part you most often forgot. For most people it is the output format or the constraints.
- Tape the template next to your monitor for a week.
The goal is not to talk more. It is to make the structure automatic, so the extra words you speak carry meaning instead of noise.
