Source: https://fnflow.app/dictate-ai-prompts

# How to dictate a good prompt for AI coding tools
A useful prompt for a coding agent is mostly context: what you want, where the code lives, what must not change and how you'll know it worked. That's tedious to type and easy to say. Speech has its own failure mode, though. You ramble, you correct yourself halfway through a sentence, and the transcript keeps every word. Here's how to speak a prompt that still reads well once it's text, whether it goes to Cursor, Claude Code or ChatGPT.

## Why talking makes prompts better and worse

When you type, you economize. The prompt shrinks to "fix the login bug", and the agent has to guess the rest. When you talk, you tend to explain why, mention what you already tried and name the file you suspect. That extra context is often exactly what the model was missing.

What you lose is structure. Spoken thoughts arrive in the order you think of them, with false starts and "actually, no, the other one". FnFlow types exactly what you said, with punctuation and capitals and no AI rewriting, so the order you speak in is the order the agent reads. A few seconds of planning before you hold the key pays off.

## An order that works out loud

Anthropic's best-practice guide for Claude Code asks for specific context. Reference the files, mention constraints and point to example patterns. For a bug, give the symptom, the likely location and what "fixed" looks like. The guide also recommends giving Claude a check it can run, such as tests or a build. It's written with typed prompts in mind, and it turns into four spoken beats without much effort:

- The goal, in one sentence. Say it first so everything after it has a frame.
- The context. Which part of the code, what you saw happen, what you already tried.
- The constraints. What has to stay as it is, such as a public API, a file nobody should touch or a dependency you can't add.
- The check. A test to write or run, a command that should pass or the output you expect.

Pause between beats. A pause costs nothing in the transcript, and it gives you a moment to decide what comes next instead of filling the gap with "so, um".

## An example

Here's a made-up prompt in that order, the way it might come out of one take:

Users get logged out after about an hour, even though their refresh tokens
are still valid. The refresh happens in the session middleware, and I think
the expiry check compares seconds with milliseconds. Don't change the cookie
format, other services read it. Write a failing test that reproduces the
early logout first, then fix it and run the auth tests.
Each sentence does one job. The first is the symptom, then come the suspect and the constraint, and the last sentence says how the agent should prove the fix.

## What to type instead of saying

Some things are faster and safer from the keyboard.

- File references. In Cursor you type @ in the chat input to attach files and folders, and Claude Code uses @ to reference files too. Dictate the prose, then type the @ references where they belong.
- Error messages and stack traces. Paste them. Reading a trace aloud invites transcription errors in the one part that has to be exact.
- Identifiers the model keeps mishearing. Add them once to Settings → Vocabulary, where FnFlow passes up to 60 terms to Whisper as spelling hints. On a test sentence, five terms in the vocabulary turned "update Zapot" into "update Saypad".

## Read it before you send it

An agent acts on what it receives. A misheard file name sends it to the wrong place, and a lost "don't" turns a constraint into an instruction. Read the prompt once before pressing Enter, and look hardest at names, numbers and negations.

Most voice inputs leave the sending to you. Cursor's docs tell you to review the transcription before sending, and Claude Code's /voice inserts the transcript and waits for Enter unless you turn on auto-submit. FnFlow pastes the text with a single ⌘V and stops there, so Enter is always your own keypress.

## Long prompts and planning sessions

For a prompt longer than a breath or two, double-tap fn to go hands-free and tap again when you're done. There's no word quota, so a long planning brain dump costs the same as a one-line request. If a paste ever goes missing, the dictation is in FnFlow's history with its audio, ready to copy again.

Tool-specific setup is covered in [voice dictation for Cursor](dictation-for-cursor) and [dictation in Claude Code and the terminal](dictation-claude-code-terminal).

Tool by tool: [vibe coding with voice input](vibe-coding-voice-input).

## Questions

**Should I clean up filler words before sending?**
Fix whatever changes the meaning first: misheard names, numbers and negations. Leftover filler makes a prompt longer but rarely changes what you asked for. FnFlow keeps your wording as you said it, without AI rewriting.

**Can I dictate code?**
Describe the code instead of reading it out. Say what a function should do, then type or paste the exact identifiers, paths and snippets. Names you use every day belong in FnFlow's vocabulary.

**Does dictating a prompt send it anywhere?**
With FnFlow, recognition runs on your Mac and the audio stays there. The prompt goes wherever Cursor, Claude Code or ChatGPT sends it after you press Enter.

**What about Cursor's microphone or Claude Code's /voice?**
Both work inside their own app. Cursor has a microphone button in the chat input. Claude Code's /voice streams audio to Anthropic's servers and needs a Claude.ai account. FnFlow works in any app and keeps the audio on your Mac.
