Voice Commands for AI Agents

Speak to agents instead of typing — useful phrases, session rules, and tools as voice becomes a primary control surface

← AI Agents  ·  AI  ·  ChatGPT Work / Codex  ·  Harness  ·  Agentic setup

Why this page exists

As agents take more of the keyboard work, your bottleneck becomes how fast and clearly you express intent. Voice is often 2–4× faster than typing for long, rambling, technical prompts — and native voice modes (Codex, ChatGPT Voice) add commands, not just dictation.

Below is a living list of phrases and operating rules collected from practitioners, with citations. Copy them into your own skill / “voice boot” doc.

Codex voice control — useful commands

Source: @ForwardEditor · “VOICE CONTROL in Codex is a GAME CHANGER” (Jul 24, 2026)

Forward Future Brian (editor for Matthew Berman’s team) shared the most useful Codex Voice Control commands he’d found after using it hands-free.

“Increase speaking speed by 10%”

Tune TTS pace so the agent keeps up with your workflow without slowing you down.

“When done with a task, don't speak your output, only say 'done.' Say your task output when I say 'go.'”

Stops long verbal dumps. You stay in control: short completion ping, full answer on demand.

“Open browser or browser tabs to (whatever you want)”

Hands-free navigation while the agent works — e.g. docs, staging, X, dashboards.

“Open the task that owns this feature”

Codex-specific: jump to the owning task/thread for the feature under discussion.

“Reorganize my windows so they fit on the screen”

Desktop hygiene by voice when multitasking agents + IDE + browser.

“Start a timer”

Lightweight timeboxing without leaving the session.

Patterns (not exact phrases)

  • Describe bugs and lag out loud: If your computer has some kind of bug or lag, literally just speak it out loud. It’ll give fixed instantly ~80% of the time (per source experience).
  • Hands-free publishing: The source opened X and posted the tip list without touching mouse or keyboard — a demo of full voice control end-to-end.

ChatGPT Voice — operating rules after night one

Source: @ForwardEditor · “After 1 night of using ChatGPT Voice” (Jul 24, 2026) — quotes the Codex list above

Same author, after a night on ChatGPT Voice: less “magic phrases,” more how you structure work so the light live model doesn’t become the bottleneck.

1

Never do work inside the live chat — only delegate

The live voice chat is set to a light model (e.g. Terra-Light). Treat it as a switchboard, not the factory floor.

2

Always delegate tasks to new threads

Be specific about what each thread owns. Use a relay skill to auto-pick model + reasoning effort for new threads, or clarify per task.

3

Expect automatic reporting to OpenAI

Be ready for Voice to automatically report your threads to the OpenAI team without asking — privacy / ops awareness, not a “feature request.”

4

Sessions are on a timer (~1 hour)

Voice will automatically end after about an hour. Plan for restarts and don’t assume an all-day session.

5

Voice resets context every start

Create a skill you can call at the start of every chat to get yourself back up to speed immediately.

6

Any noise can nuke the prompt

Voice still does the annoying thing where if you make any noise at all you might lose your entire previous prompt — keep a short written backup or re-speak carefully.

Suggested “boot” skill (implement yourself): At the start of every Voice session, say something like “Run my voice boot skill” — restore project context, open constraints, active threads, and preferred models — because Voice resets context every start.

Quick cheat sheet

Say / do Why Where
“Only say done… output when I say go” Short feedback loop; no monologues Codex Voice
“Increase speaking speed by 10%” Match your thinking pace Codex Voice
“Open the task that owns this feature” Jump to owning work unit Codex
“Reorganize my windows…” / “Start a timer” Environment + timeboxing by voice Codex Voice
Delegate only; new thread per task Live chat is light model / switchboard ChatGPT Voice
Call a boot skill every session start Context resets; sessions ~1 hour ChatGPT Voice
Describe the bug / lag out loud Fast diagnosis + fix path Codex Voice

Why voice beats typing for agents

Once agents handle most of the work, your main job is expressing intent clearly and quickly. Humans speak ~120–150 wpm vs typing ~40–60 wpm. For long natural-language prompts (“try decreasing the learning-rate scheduler warmup by 20% but increase batch size if memory allows…”), voice is 2–4× faster and keeps flow.

  • Agents (Claude Code, Codex, Cursor, etc.) thrive on detailed conversational prompts.
  • Speaking lets you “vibe” edge cases and aesthetics without breaking to type.
  • Error handling: read a traceback aloud instead of copy-paste gymnastics.
  • Native voice modes go further: control windows, threads, and completion style by phrase.

Dictation tools (when the agent isn’t native voice)

If you’re in a terminal/editor without Codex/ChatGPT Voice, use a hotkey dictation stack — same “speak intent” muscle, different pipe.

SuperWhisper

Mac-focused (also Windows/iOS), offline-capable. Strong on technical terms; hotkey anywhere.

superwhisper.com →

Wispr Flow

Cross-platform; hold key, speak → text. Solid with Cursor / Claude Code workflows.

wisprflow.ai →

WhisperTyping

Whisper-powered dictation in any app; free tier available.

whispertyping.com →

MacWhisper + VS Code Whisper Assistant

Local Whisper for privacy; editor hotkey into the agent composer.

MacWhisper →

Common flow: hotkey → speak → auto-paste into agent chat → accept changes.

How this fits the stack

  • ChatGPT Work / Codex — models, Work vs Codex tabs, computer-use style workflows.
  • Agentic setup checklist — boot skills, worksheets, and self-healing docs pair with “voice boot every session.”
  • Harness — practical agent coding; voice is another input into the same harness.
  • Headless Empire OS — keep TODAY / constraints written so voice sessions don’t only optimize busywork.
  • Hormozi on AI leverage — speed of input ≠ money unless aimed at the real constraint.

Sources

Comments

Approved comments appear below. Log in once with GFAVIP — it applies across the whole site. GFAVIP login

View comments archive