On this page
Back in March, I was wishing coding and CLI AI tools had built-in voice input. Tools like Claude Code and Codex eventually rolled it out, but once microphone support arrived, that turned out to be the easy part. The real challenge was figuring out how to express my thoughts out loud without breaking my stride.
At first, I assumed I would mostly give short instructions, so I set up push-to-talk tools like SayIt where you hold a key to record and release to transcribe. Over time, though, whenever I needed to walk through a curriculum deck, review system architecture, or run an end-to-end acceptance check, I gravitated toward long, uninterrupted recordings. Push-to-talk handles quick chores; long recordings are for deep thinking where interruptions hurt. I am still balancing the two.
See the 3-month 1-on-1 program →
Two Voice Prompting Modes: Quick Push-to-Talk vs. Long Audio for Deep Work
The first mode is push-to-talk for small, routine tasks. Just like SayIt, you press to record and release to transcribe. For quick chores, it works well.
When reviewing slide decks, course structures, system architectures, or end-to-end deliverables, my habit shifted: open QuickTime, talk through feedback while browsing the material, and then drop the audio into a transcription tool for an AI agent to execute.
Whenever you need to untangle complex ideas without breaking flow, record the whole session. Once you are done, pass the audio file to your transcription setup to generate a transcript and trigger downstream tasks.
Transcribing Long Audio: Zero-Setup Web UI and Headless Multimodal Calls
With long recordings, the bottleneck is transcription.
To chew through long audio quickly, the best approach is leaning directly on multimodal models. There are two straightforward paths:
The simplest, zero-setup option is dropping the audio file straight into the Gemini web UI. It is hard to beat: drop a 23-minute recording into the chat, and Google’s compute churns out a clean, typo-free transcript in 55 seconds. Pair this with your other AI agents: record long audio while reviewing documents to stay in flow, then copy the transcript over and let the agent get to work.
If you want transcription baked into your workflow without switching to a browser, you can call a terminal tool in headless mode directly from Claude Code or Codex. The key here is letting the multimodal model use its native audio understanding—feeding the raw audio directly for it to listen and transcribe, rather than piping it through Whisper or slicing it up with ffmpeg first. Native multimodal transcription matches traditional tools in accuracy, runs significantly faster, and includes speaker diarization out of the box.
Also, if your Google Drive plan comes with bundled AI credits, burning that quota via headless CLI calls from Codex or Claude Code for grunt work—voice transcription, image recognition, document summaries—is a great deal. The quota comes bundled with Drive anyway; just the other day I spun up 10 Gemini agents simultaneously to read a 193-page report without feeling the pinch.
Practical Tips and Guidelines
A few practical takeaways when integrating long audio into your daily routine:
First, filler words. Verbal fillers like “um” and “uh” can be cleaned up afterward if you need polished text, but if you are just prompting an agent, you can ignore them. On a subscription quota, a few extra tokens make no difference.
Second, alternative STT models. People often ask how accurate this approach is. The error rate is comparable to standard tools, but it runs much faster and handles speaker identification automatically. Others have recommended alternative speech-to-text models; I tested one that excelled at Taiwanese Hokkien, though its Mandarin accuracy fell slightly behind Qwen.
Third, prerequisites. Make sure you are logged in via the antigravity CLI beforehand.
Appendix
For anyone looking to put their Google Drive AI credits to good use, you can have Claude Code or Codex set up a skill using this configuration:
agy -p "Transcribe the following audio file verbatim using your native multimodal audio understanding. Do not call whisper, ffmpeg analysis, or any external speech-to-text tools; listen to and transcribe the audio file directly. Audio path: <absolute path>. Transcribe verbatim without summarizing or omitting anything, and label the speaker and timestamp [MM:SS] at the start of each section." --model "gemini-3.6-flash-high" --add-dir <directory containing audio file> --dangerously-skip-permissions
The core of this command lies in prompt design: forcing the model to rely on its native audio comprehension, bypassing external Whisper calls and ffmpeg preprocessing, and returning a full verbatim transcript with speaker labels and [MM:SS] timestamps, pinned to gemini-3.6-flash-high.