Dictate to AI Coding Assistants: Cursor, Claude Code & Copilot
Speaking prompts beats typing them for AI-assisted coding. How to dictate to Cursor, Claude Code, and Copilot Chat with a system-wide hotkey that types into any field.
AI coding assistants flipped the bottleneck. The hard part is no longer writing the loop or remembering the exact API signature; it's describing what you want clearly enough for the model to do it right the first time. And description is a language task, which means the fastest input device you own is your mouth, not your hands.
Bottom line: prompting an AI assistant is talking, not typing. People speak at roughly 150 words per minute and type at 40–55. If you dictate your prompts and keep the keyboard for navigation and fine edits, you describe richer intent in a third of the time — across the terminal, the chat panel, and the PR box alike.
Why prompts are the new bottleneck
Studies of programmer typing put working speed around 40–55 words per minute, while comfortable speaking lands between 120 and 150 — about a 3x gap. For boilerplate that gap never mattered, because you were thinking as fast as you typed. Prompting is different. A good prompt names files, explains an edge case, states the constraint, and gives an example. That is exactly the kind of nuanced, run-on natural language that's painful to type and effortless to say.
The win isn't just speed; it's completeness. When typing is cheap you write "add retry to the fetch." When speaking is cheap you say "wrap the fetch in lib/api/client.ts in a retry with exponential backoff, cap it at three attempts, and don't retry on 4xx." The second prompt gets you working code; the first gets you a round trip.
The catch: native voice only covers one box
Each assistant has added some voice, but the coverage is uneven and the boundaries matter.
- Cursor shipped a native voice mode (the mic icon in the chat input) in version 2.0. It's genuinely handy — but it only dictates into Cursor's own chat panel. The moment you tab to the integrated terminal, a Git commit box, a browser, or Slack, the mic goes quiet.
- Claude Code added a built-in
/voicecommand (v2.1.69 and later) with hold-to-record and tap-to-toggle right inside the CLI. Good news if you live in that one terminal; note that its audio is streamed to the vendor's servers for transcription rather than processed on-device. - GitHub Copilot has no first-party system-wide dictation. Copilot Chat takes typed prompts; voice is whatever your OS or a third-party tool provides.
As one developer put it, a microphone that only works in a single chat panel covers maybe 20% of where you actually type prompts. The other 80% — terminal commands, PR descriptions, code review comments, the issue you're filing, the Slack message asking a teammate — is left out.
Honest caveat: if you spend literally all day inside Cursor's chat panel and nowhere else, its built-in mic may be all you need. The case for a system-wide tool is strongest when your day spans a terminal, a browser, a chat app, and an editor — which is most real workflows.
One hotkey, every field
Pithflow takes the opposite approach to a per-app mic: it's a Windows-native dictation layer that types into whatever field has focus. The flow is muscle memory after a day:
- Hold the hotkey (Ctrl+Space by default).
- Describe the logic out loud — the task, the file, the edge case.
- Release. A cleaned-up, punctuated prompt is typed into the active input.
Because it operates at the OS level, the same hotkey works in the Claude Code terminal, Copilot Chat, Cursor's composer, a GitHub PR description, a JIRA ticket, VS Code, Word, and your browser. It also works over Citrix, RDP, and other VDI sessions, where editor-embedded voice features usually can't reach. There's a hands-free tap-to-toggle for longer dictation (up to 20 minutes) when you're narrating a whole spec.
Where voice wins and where the keyboard still wins
Voice is for intent; the keyboard is for precision. Don't try to dictate a regex character by character — say "match a US phone number" and let the assistant write it. The split that works in practice:
| Task | Best input | Why |
|---|---|---|
| Describing a feature or refactor | Voice | Long natural language, fast to say, carries nuance |
| Commit messages & PR descriptions | Voice | Prose, not code — ideal dictation territory |
| Code review comments | Voice | You're explaining reasoning, not editing syntax |
| Cursor navigation, selection, scrolling | Keyboard / mouse | Spatial actions are faster by hand |
| Renaming a variable, fixing one character | Keyboard | Surgical edits beat a voice round-trip |
| Exact symbols / punctuation-heavy strings | Keyboard | Faster to type than to spell aloud |
Making technical terms transcribe correctly
The number one reason developers abandon dictation is that the transcriber turns useEffect into "use effect," npx into "NPX," and --no-verify into "no verify." Pithflow handles this two ways:
- Specialty term packs — including an engineering pack — bias recognition toward programming vocabulary so library names, CLI commands, and common API terms land correctly out of the box.
- A personal dictionary where you add the names that are specific to your stack: your internal service names, an unusual library, your company's acronyms, a teammate's GitHub handle. Add it once and it transcribes right every time.
You can also save snippets — voice-triggered text expansion — for boilerplate you dictate constantly, like a standard PR checklist or your conventional-commit prefix.
What it costs against the alternatives
You're already paying for the assistant; dictation is a small add-on. For honesty, here's where things actually stand in mid-2026. Several dictation tools are excellent — if you're on a Mac, you have more options than you do on Windows, and we'll say so plainly.
| Tool | Price | Scope | Platform |
|---|---|---|---|
| Pithflow | Free (2,000 words/wk); Pro $9.99/mo or $99/yr | System-wide, any field | Windows |
| Cursor native voice | Included with Cursor Pro ($20/mo) | Cursor chat panel only | Cross-platform |
Claude Code /voice | Included (Claude Code is usage-based) | Claude Code CLI only | Cross-platform |
| Wispr Flow | $15/mo ($12/mo annual); free 2,000 words/wk | System-wide | Mac, Windows, Android |
| Aqua Voice | $8/mo ($96/yr); free 1,000 words | System-wide | Mac-first, iOS |
For the AI assistants themselves, GitHub Copilot Pro is $10/month, Cursor Pro is $20/month, and Claude Code bills by usage (heavy sprints can hit $50–100). If you're choosing between dictation tools, the honest read: Aqua Voice is cheaper and very accurate but Mac-leaning; Wispr Flow is the broadest cross-platform option at a higher price. Pithflow's pitch is specifically the Windows developer who wants the lowest paid tier and a true free plan with no card required. Compare them side by side on our comparisons page before you commit, and read more workflow breakdowns on the blog.
A realistic day with dictated prompts
Morning: you open Claude Code in your terminal, hold the hotkey, and describe the migration you want — three sentences of context you'd never have bothered typing. You read the diff, click into the file, and fix one import by hand. You write the commit message by voice. You open the PR in your browser and dictate the description, then drop a review comment on a colleague's PR explaining why the lock should be narrower. Not once did you switch tools or hunt for a mic icon — the same hotkey carried you through the terminal, the editor, and the browser. That continuity is the whole point.
FAQ
Doesn't Cursor already have voice built in?
Yes, since version 2.0 — but only inside its chat panel. It can't dictate into the integrated terminal, a commit box, your browser, or another app. A system-wide tool like Pithflow types into any focused field, so you're not limited to one window. If you genuinely never leave Cursor's chat, the native mic may be enough.
Can I dictate prompts into the Claude Code terminal?
Two ways. Claude Code added a native /voice command (v2.1.69+) with hold- and tap-to-record, which streams audio to the vendor's servers. Or use a system-wide tool that types into the terminal like any other field — the same hotkey that works in your editor and browser also works at the prompt, with no per-app setup and your dictionary applied everywhere.
Will it get my library names and CLI flags right?
Plain dictation usually mangles them. Pithflow's engineering term pack biases recognition toward programming vocabulary, and the personal dictionary lets you add anything specific to your stack — internal service names, unusual libraries, acronyms — so they transcribe correctly every time. For exact punctuation-heavy strings, it's still faster to type.
Is my audio or code stored anywhere?
With Pithflow, audio is processed in real time and never stored on our servers; your session is encrypted on-device using Windows DPAPI. Note that some editor-embedded voice features, including Claude Code's /voice, send audio to their vendor's servers for transcription — worth checking your own tool's policy if that matters to you.
Is there a free way to try this?
Yes. Pithflow has a free tier of 2,000 words per week with no credit card, which is plenty to test the workflow across Cursor, Claude Code, and Copilot before deciding. Pro is $9.99/month or $99/year, and there's a 7-day trial — see pricing for the current details. Note Pithflow is Windows-only today; a Mac build isn't available, though an Android keyboard is in development.
Try Pithflow free
Voice dictation that's faster than typing. Hold a key, speak, get clean text in any Windows or Mac app. Free tier: 2,000 words a week, no credit card.