Coding with AI and Voice Commands: A Forced Experiment with Unexpected Benefits

A few months ago, I broke my hand while biking to work in San Francisco, leaving me capable of typing only with my left hand. This unexpected setback forced me to rely heavily on a combination of voice-to-text and Claude (Anthropic's AI assistant) for coding. The results were surprising: not only did I manage to stay productive, but I also wrote over 3,000 lines of code in one week-mostly boilerplate, but still impressive given the circumstances.

This experience has convinced me that AI-assisted coding is not just a future possibility but a present reality with significant benefits. Here’s how I set up my workflow and what I learned along the way.

My Setup

Voice-to-Text on Mac

Initially, I tried using the built-in Voice Control feature on macOS. While it was decent for natural language tasks, it struggled with coding-specific commands. Terms like "eval" were often misinterpreted as "evil," making it impractical for writing code directly.

Claude and Voice Commands

Since I was already familiar with AI code generation tools like Copilot, I decided to leverage Claude, Anthropic's AI assistant. My setup involved copying large chunks of the codebase into Claude and giving voice commands to transform or generate code. For example:

• "Refactor ABC function to take input XYZ"
• "Write unit tests for these new functions ABC, using examples from XYZ"

Claude wasn't perfect on the first try, but it was receptive to follow-up instructions and tweaks. This felt like pair programming with an AI partner who could execute my verbal commands.

Copilot Autocomplete

While Copilot is a powerful tool, it became too slow for me in this context. It required me to write out half a line of code before suggesting completions, which was cumbersome with one hand. Expressing my goals in English and using Claude's capabilities proved much more efficient.

How to Work Effectively with Claude

Be Specific and Give Examples

One of the key lessons I learned is the importance of being specific and providing examples when giving instructions to Claude. Generic requests often result in generic, non-specific code that doesn't fit your project. Here are some tips:

• Explicit Instructions: Clearly state what input and output you expect, which libraries to use, and any other relevant context.
• Contextual Placement: Place your instructions at the beginning and end of your input data to ensure Claude understands the full scope of your request.
• Examples: Provide examples of similar code or desired outcomes. This helps Claude generate more accurate and useful results.

Managing Output Length

Claude has output length limits, which can be frustrating when working with large codebases. I found myself frequently copying and pasting between my IDE and Claude, manually stitching together truncated code snippets. While this was sometimes tedious, it was manageable once I got into a rhythm.

Handling Frustrations

There were moments when Claude would forget earlier instructions or produce unexpected results, leading to some frustrating interactions. However, the ability to give follow-up commands and make tweaks kept the process productive. Being able to vent my frustrations verbally (thanks to voice-to-text) also made it more satisfying!

The Future of Coding

This experience has opened my eyes to the potential of AI-assisted coding. While there are still challenges, such as output length limits and the need for specific instructions, the benefits are undeniable. Pairing voice commands with powerful AI tools like Claude can significantly enhance productivity and make coding more accessible.

For anyone considering a similar setup, I highly recommend experimenting with AI assistants and voice-to-text systems. You might be surprised by how much you can achieve.