The short answer: in ChatGPT, Claude, Gemini, Cursor, Codex, or Claude Code, click the prompt field, press ⌃+⌥+R (Mac) or Ctrl+Alt+R (Windows/Linux), and dictate a full task spec with repo context, constraints, acceptance criteria, and expected output format. AICHE inserts cleaned text at the cursor. You review and send.
Typed prompts shrink under keyboard friction. You skip the error trace, the file path, and the verification command. Spoken prompts carry what agents need: system context, reproduction steps, constraints, and acceptance criteria in one pass.
Example (dictated into Cursor): "Repo acme/web, branch fix/oauth-callback. Bug: after Google OAuth, user lands on / instead of /dashboard. Check src/auth/AuthCallback.tsx and localStorage key returnUrl. Acceptance criteria: redirect respects stored return URL, add test in authCallback.test.ts, run pnpm test authCallback. Output: patch summary plus test command results."
How It Works
- Open the AI surface: Claude web, ChatGPT, Gemini, Cursor agent chat, Codex, or Claude Code terminal input.
- Click in the prompt or message field (where typed text would go).
- Press ⌃+⌥+R (Mac) or Ctrl+Alt+R (Windows/Linux) once to start recording (toggle, not push-to-talk).
- Speak the prompt: context first, then the ask, then constraints and acceptance criteria.
- Press the hotkey again. AICHE transcribes, runs Message Ready cleanup, and inserts at the cursor.
- Read the prompt, edit names or paths if needed, then send or press Enter.
AICHE does not run the agent, approve diffs, or control the IDE. It only supplies prompt text.
Why Spoken Prompts Get Better AI Responses
Short typed prompts look like: "fix auth redirect." A dictated prompt names the stack, symptom, files, and done condition:
"I'm on Next.js 14 App Router. After OAuth, AuthCallback reads returnUrl from localStorage but the key is cleared on login. Trace redirect in AuthCallback.tsx, preserve returnUrl through the callback, and add a test that covers dashboard return. Do not change the public API of useSession."
That is the difference between a guess and a scoped task. Speaking for 30-60 seconds routinely yields 150-250 words of context most people will not type.
Best Practices for Voice Prompting
Give Context First, Then the Ask
Start by explaining what you're working on, then describe the specific problem, then state what you want the AI to do. This mirrors how you'd explain something to a colleague, and AI assistants respond best to this structure.
Don't Self-Edit While Speaking
When typing, you constantly backspace and rephrase. When speaking, resist that urge. Let your thoughts flow out in order - Message Ready cleans up the grammar and removes filler words afterward. Your job is to communicate the idea, not to craft perfect sentences in real-time.
Use 30-60 Second Recordings
This is the sweet spot. Under 15 seconds and you're probably leaving out important context. Over 90 seconds and you're likely wandering. If your prompt needs more detail, send it in two recordings - one for context, one for the specific ask.
Walk While You Dictate
This isn't just a productivity trick - research on embodied cognition shows that physical movement genuinely improves problem-solving. Walking while you speak your prompt helps you think through the problem more completely than sitting and typing.
Works With Every AI Tool
AICHE's hotkey is system-wide, so it works in any application with a text input:
- Claude (web and API) - dictate complex reasoning prompts
- ChatGPT - speak questions and follow-ups naturally
- Cursor - dictate code change requests with full context
- GitHub Copilot Chat - describe the code you need
- Perplexity - ask research questions in natural language
- Any AI chat interface - if it has a text field, AICHE works with it
Multilingual Prompting
If you think in a non-English language, enable Translate to English in AICHE settings before recording. You speak naturally in your native language, and the output converts to English automatically. This is especially useful for non-native English speakers who can articulate complex technical problems more precisely in their first language.
The Numbers
A typical AI-heavy workday involves 30-50 prompts. At 3-5 minutes of typing per detailed prompt, that's 90-250 minutes of typing daily. With voice dictation at 30-60 seconds per prompt, the same work takes 15-50 minutes - and the prompts contain more context, so you get better responses with fewer back-and-forth cycles.
Do this now: open Claude or ChatGPT, press your hotkey, and dictate one real coding problem you're working on today. Don't think about how it sounds - just explain the problem the way you'd tell a colleague.