Voice shines with AI assistants because good prompts are long and typing them is a tax. These patterns work in any chat or agent tool.
The context dump
"I am trying to get the login form to validate on blur. I already tried adding an onBlur handler and it fires twice. Look at the form component first and tell me why before changing anything."
Goal, what you tried, where to look, in that order. Thirty seconds of talking replaces five minutes of typing the same context.
The constraint sandwich
"Write a product update for the beta group. Constraints: under 150 words, no exclamation marks, plain language. Output as a single paragraph I can paste."
Goal, constraints, output format. Saying constraints out loud is where dictation beats typing: you would have trimmed them to save keystrokes.
Rubber duck debugging, recorded
"Here is what confuses me. The request succeeds in curl but fails from the app with the same headers, which should be impossible, unless the client rewrites something."
Explaining a bug aloud to the AI often solves it before the answer comes back. The transcript doubles as your notes.
Structure my ramble
"I am going to think out loud about our onboarding problem. Turn what I say into a short list of the distinct problems I mention, ignore the repetition."
Tell the model up front that a ramble is coming and what to do with it. Then genuinely ramble.
Meeting to action items
"From what I am about to say, extract action items with an owner and a date each. We agreed Sara ships the pricing page Friday, I review the copy Thursday night, and legal sees it before it goes live."
Dictate from memory right after the call while it is fresh. The model formats; you just talk.
Steelman both sides
"Give me the three strongest arguments for rewriting in Swift and the three strongest against, then say which side you find stronger and why."
A decision prompt you will use weekly. Say the options; do not prejudge them in your phrasing.
Read the error aloud
"The build fails with: cannot find type Publisher in scope, in AudioEngine line 40. It started after I removed the Combine import. What is the minimal fix?"
Reading an error message is faster than copying it when you are in a terminal or a screenshot. Include where and when it started.
The spoken spec
"New setting: retention. As a user I want dictation history deleted automatically after a period I choose, including audio, so old recordings never pile up. Default 30 days. Write acceptance criteria."
User story plus one default. The model expands it; you stay at the level of intent.
Steer without retyping
"Keep the structure and the facts, change the tone to something a tired engineer would actually read, and cut it by a third."
Follow-up steering is the highest value per word you will ever dictate. Name what stays, then what changes.
The background briefing
"Before you answer, background: we are a two person team, the app is a paid Mac utility, we do not want a server. Now, how should we do license validation?"
Say the context you always retype. Better: dictate it once into a note and reuse, appending the fresh question by voice.