Hands-Free Coding: Dictate Prompts to Claude Code from Your iPhone

The prompt is the bottleneck.

Not the agent, not the network, not the server — the prompt. Specifically, the moment when you need to write a careful, detailed instruction on a 6-inch glass keyboard while standing in a hallway between meetings.

That friction is real. And it quietly discourages you from using your phone to steer AI coding sessions the way you otherwise could.

Tactic Remote includes voice input to address this directly. You can compose prompts by speaking, review the transcription, and send — without typing.

Why long prompts are painful to type on a phone

Most useful AI coding instructions are not short.

"Fix the bug" rarely works. The prompt that actually moves the work forward looks more like this:

"In the AuthService class, the token refresh logic is failing silently when the network times out. Check the catch block in refreshToken(), compare with how we handle it in SessionManager, and add a retry with exponential backoff. Don't change the public interface."

That is 45 words. On a desktop keyboard, that takes seconds. On a phone, it takes meaningful effort: autocorrect fights you, class names and technical terms get mangled, and switching between character sets slows you down further.

The result is predictable: you shorten the prompt, lose precision, and the agent does something adjacent to what you wanted rather than exactly what you needed. Or you just wait until you're back at a keyboard.

Both outcomes represent waste. The agent is running. The session is live. The moment is there. But the input surface is in the way.

What voice input does in Tactic Remote

Tactic Remote lets you compose prompts using Speech-to-Text.

The flow is simple:

1. Tap the microphone icon in the prompt area.
2. Speak your instruction at a normal conversational pace.
3. The transcription appears as editable text in the prompt field.
4. Review and correct it if needed, then send.

You are not sending raw audio to the agent. The voice input produces text — the same text you would have typed. You remain in control of what gets sent before it leaves your phone.

This matters for a few reasons. Speaking is faster than typing for most people, especially for long, structured instructions. The review step keeps you accurate: you catch technical term misrecognitions before they reach the agent. And the result integrates naturally with the rest of Tactic Remote's prompt flow, including Quick Prompts and per-session agent selection.

When voice input is actually useful

Voice input is not equally valuable in every situation. Here is where it makes the most difference.

Long natural-language instructions. When you need to explain context, reasoning, or intent — not just a command — speaking is faster and more natural than typing. Prose flows out of speech; it fights the phone keyboard.

While walking or commuting. The most common mobile AI coding pattern is checking in on a running session while away from a desk. If the agent has finished a task and needs direction, being able to speak the next instruction rather than pause, sit down, and type it keeps the session moving.

Technical terms in context. Saying "refactor the OAuth flow in UserSessionController" is awkward to type on mobile — autocorrect will attempt to ruin it. Speaking the full phrase and then reviewing the transcription is often cleaner than fighting the software keyboard.

Accessibility. For developers who find typing on small screens difficult due to motor or visual constraints, voice input reduces a real barrier to using mobile AI coding tools.

First drafts of complex instructions. Sometimes you know what you want but need to think out loud to structure it. Speaking a first draft, reviewing the text, and editing before sending can produce a cleaner final prompt than typing from scratch.

How to use it well

Voice input is a composition tool, not a dictation pipe. The review step is part of the workflow, not a formality. A few practices help.

Speak in complete sentences with structure. "Check the error handling in the login flow, compare it with the signup flow, and list any inconsistencies" works better than "um, like, check the login error handling I guess." The more structured your speech, the less cleanup the transcription needs.

Name things precisely. File names, class names, function names — say them clearly and completely. After transcription, scan those terms first. If "AuthService" became "Awesome," correct it before sending.

Review before you send — always. The review step is what makes voice input reliable for technical work. Spend two seconds scanning the transcription for misrecognized terms. This is faster than fixing a confused agent response downstream.

Combine with Quick Prompts for repeated workflows. Voice input is best for instructions that are context-specific and long. For prompts you send repeatedly — "review the current diff," "summarize this session," "add tests for the changed behavior" — Quick Prompts are the right tool. Use them together: reach for voice when the instruction is unique to this moment, reach for Quick Prompts when the pattern is one you've used before.

Edit the transcription freely. The voice input field is just a text field. After speaking, you can tap anywhere, correct misrecognized words, add technical details that are easier to type than say (a specific line number, a URL, a flag name), and adjust phrasing. Treat it as a starting point, not a final output.

A realistic example

Here is what a voice-input workflow looks like for a refactor instruction.

You are on the train. Your agent is running on your Mac at home. You got a push notification that it finished the previous task and is waiting. You open Tactic Remote, see the completed output, and need to give the next instruction.

You tap the microphone and say:

"In the DataSyncService, the full sync method is running on the main thread. Move it to a background queue and make sure the completion callback is dispatched back to main. Check if there are any existing tests for this and update them if needed. Don't touch the partial sync method."

The transcription lands in the prompt field. You scan it: "DataSyncService" is correct, "background queue" is correct, "dispatched back to main" is correct. You send.

That instruction — 61 words — took about 15 seconds to speak and 5 to review. Typing it with one thumb on a subway would have taken considerably longer, and the technical terms would have required careful correction.

The agent is now unblocked. You put the phone in your pocket.

Limits and realistic expectations

Voice input is useful, but it is not magic.

Technical jargon — especially invented class names, project-specific terms, or unusual identifiers — will sometimes be misrecognized. The review step catches most of these before they cause problems.

Background noise affects transcription quality. A quiet environment produces cleaner results than a busy street or a noisy office. If accuracy drops, shift to typing or find a quieter moment.

Very long instructions (several paragraphs) work better broken into focused segments. Speaking one clear instruction at a time is more reliable than trying to dictate an entire multi-step plan in one breath.

The voice input does not interpret or execute anything — it produces text. What you review in the prompt field is exactly what the agent will receive.

Prompt composition, not prompt replacement

Voice input does not change what a good prompt looks like. It changes how fast and comfortably you can write one on a phone.

The agent still needs clear context, scoped instructions, and a stated goal. Speaking a vague prompt gets you a vague response, just faster. The skill of writing effective AI coding prompts applies equally whether you typed the text or spoke it.

What voice input removes is one of the main reasons mobile AI coding feels laborious: the physical friction of composing careful instructions on a small keyboard in a non-desk setting.

That friction is not inherent to mobile. It is a tooling problem. Removing it makes the phone a more capable control surface for the agent running on your desk.

---

If you want to build out a library of prompts you reach for repeatedly, read Quick Prompts Are Workflow Memory. For a broader look at the async workflow pattern — letting the agent run, getting notified, and steering remotely — see Your AI Coding Assistant Doesn't Need You to Watch. And if you are setting up Tactic Remote for the first time, How to Run Claude Code from Your iPhone covers the full connection flow.