How Theo Works With One Hand: Dictation and Coding Agents

A hand injury has left me in and out of a cast and made typing, shortcuts and navigation difficult. Although I used to type quickly, I have had to change how I work. This is a personal account of those adjustments, as well as a collection of ideas that might help other people who want to interact with a computer differently.

‘Coding without typing’ is really a story about staying productive without typing as much. Dictating the exact syntax of a function is awkward for me. Voice transcription and coding agents are useful because I can describe what I want and let the software work out the implementation.

The problem extends beyond writing code: switching applications and using keyboard shortcuts are uncomfortable with one hand. One small change is to leave edges of two windows visible rather than make both full screen, so I can switch with the mouse. I look for ways to make the remaining hand do less work.

My phone is easier for me to use one-handed because I have long used swipe typing. I have also invested in mobile interfaces for the tools I use, including T3 Code, so I can start or check work from the phone. Accessibility needs and the best input method will differ from person to person.

Voice dictation lets me get thoughts onto the screen, but it does not by itself switch apps or navigate the computer. I also felt uncomfortable speaking to my laptop around coworkers. That social friction mattered even though I talk on camera all the time.

A small close microphone changed that for me. I can speak much more quietly than I expected, and in my demonstration the dictation tool still understands me. I suggest trying the built-in laptop microphone first at a normal volume; the separate microphone helped me most with whispering at my desk.

The microphone cost about $70 when I bought it, and I find the combination handles office background noise well. In one demonstration, a video plays nearby while I dictate. That is my experience with this particular microphone, software and environment, not a guarantee of noise rejection elsewhere.

I use Whisper Flow and have complaints as well as praise. Ben Davis has had good results with a local alternative, and other tools may work too. The change that mattered to me was being able to dictate without feeling that I was disturbing everyone around me.

Using a terminal by voice has been particularly unpleasant for me, so I have largely stopped using it directly. Instead, I use agents and graphical interfaces for many of the operations I previously typed. That is a preference shaped by my injury and tools, not a rule that terminals are unsuitable for everyone.

One important project on my computer is called fleet. It documents my machines, what they are for, how to connect to them and what is installed. This gives an agent the context to carry out a task across computers without my manually entering every command.

For example, I was downloading a game file on one computer and wanted it on another. I told the agent where the file would appear and where to copy it, even before the download had finished, then left it to wait and transfer it. It notified me when the task was done.

Dictation still makes mistakes: it confused the names Lakebed and Leftbook, so I corrected the destination. Describing the destination and allowing the agent to wait moved some of the preparation and follow-through out of my hands. Those details still need checking, especially when a mistaken destination matters.

Another example is a development build of T3 Code’s experimental orchestrator. I wanted to try a branch without disrupting my nightly installation or its database. Rather than manually locate the branch, clone it and build it, I asked an agent to handle those steps.

I dictated the request from my phone: build the experimental branch as a separately named app, use a separate home directory and install it without interfering with the existing app. I checked it after getting ready for my stream and found the custom build in place. The explicit isolation requirement mattered to the task.

For a personal example, I needed to move medical records while changing doctors. I asked a computer-use agent to download the records from the hospital portal. I say it took about 20 minutes and involved 49 PDFs, with repeated navigation, downloading, verification and organization.

When my doctor sent the upload destination, I asked another thread to upload the files. This was my experience with my own records and my chosen tools. The convenience of delegating navigation does not remove the need to consider access, privacy and whether the destination and files are correct.

With one hand, ordinary navigation can be demoralizingly slow. Running several agent tasks in parallel has changed how I schedule work: a task no longer always needs a continuous hour of my own attention. I can start it, do something else and return to inspect its result.

I picture work as a path from an idea to a finished result. Previously I brought agents into individual portions and did a lot of planning and testing myself. Gradually I moved them earlier in that path and asked them to continue further, while still deciding what result I wanted.

I now ask agents to verify their changes with computer use, review tools and additional passes before reporting back. I may provide a problem rather than a detailed solution and ask for options or a PR. These checks are useful evidence, but they cannot guarantee that a change has no regression.

In some of my projects I also authorize agents to merge. I report more than 100 merges by Astra and at least 50 by Fable, with two noticed regressions involving removed animations. This is an anecdotal account from my projects, not an independently audited comparison with human developers or a general failure rate.

My broader goal is to involve the agent earlier and have it work longer before requesting attention. That does not make speed irrelevant: urgent bug fixes and interactive design iteration still benefit from fast responses. For background tasks I check less often, elapsed time can matter less to me.

I describe an iOS app thread where faster iteration is valuable because I have kept a great deal of work in one long conversation. Other tasks, such as background capacity work or experimentation, can run much more slowly without disrupting my day. Different tasks have different latency needs.

In T3 Code I can send a task into a new thread while staying in the composer, then start another. This makes it easier to get several ideas out of my head. The dictation demo also produces an error, which is a good reminder that speech recognition still needs attention when accuracy matters.

I care about grammar and used to correct every dictated error. I have found that models often understand a prompt despite a wrong word or typo, so I am trying to tolerate more mistakes and correct them if they cause confusion. That is a personal tradeoff; names, numbers and important instructions may need careful checking.

I keep looking for the points where the computer frustrates me. Browsing numerous GitHub tabs is one example. Instead, I ask an agent to summarize recent relevant PRs, their status and why I might care, and then use T3 Code’s in-app PR viewer to inspect changes without switching applications.

Copying context between agent threads is also awkward one-handed. Sometimes I ask a new agent to find a recent thread or relevant information itself rather than manually copy it. That can take longer and cost more tokens, but it reduces my physical interaction and lets me spend attention elsewhere.

My exercise is to try describing a problem to an agent before doing all the preparation yourself, while continuing your usual approach separately. Compare what the agent finds with your own result. Keep permissions and review appropriate for the work; the point is to learn which steps can actually be delegated.

My hand forced me to rethink both ends of the process: when to bring an agent in and when I need to return. I miss typing, but dictation, a small microphone and these workflows have made using my computer more pleasant. I hope the ideas help people adapt them to their own needs.