I used to dream about coding on the train. Now I'm managing agents from my phone. 🚉🍁🍂

Kaia Walters — Mon, 02 Nov 2026 13:50:38 +0000

This weekend I had the opportunity to work with v0, Claude Projects, Claude Code, agent skills, GitHub for Android, and PRDs to build and manage a web app from my phone.

The skill bucket list continues! Next up: leveraging agents to tackle tickets from the train. 🚉

If you're from Boston, you've probably already dreamed about being able to commute and code.

Back when I first started out in tech, my coworkers and I used to talk about having a desktop that was always running. You could remote into it and complete project tasks from a tablet. That was a fantasy for me back in the day.

Yesterday, that dream felt a little more real when I saw my first agent-generated PR sitting in the review column.

I had assigned a task, stepped away from my computer, and came back to find a pull request waiting for me. Pretty exciting!

Here's a little more about how I got there.

1. Getting the project organized

I'm volunteering to build a new web app for a church, and the inception process started off pretty much how you'd expect.

My goal was to consolidate the design research and notes from past conversations with the organization into a folder in Google Drive. I also included screenshots of example websites that I wanted to reference during the design process.

I created a second folder to outline the required features for the application and the project stack.

From there, I used that information to craft a prompt for v0 to create some potential designs for the project. I wanted to explore a few different directions, varying the color palette, fonts, and image placement to figure out which design would be the best starting point for the site.

v0 is a very powerful tool, but on the free plan you have a limited number of credits to work with. I've found that being very clear about the project stack and scoping code generation to the client side is the best way to ensure I'm ending the session with a design I can actually build on.

I also wanted to keep the initial code relatively simple and establish a monorepo that I could continue developing as the project grew. In my experience, v0 can lose its way when generating more complex backend logic, so I prefer to handle that separately.

What happened next is where I got to change my process!

2. Creating a home for project context with Claude

Creating a new Claude Project was the first step.

This is where I uploaded my research folder so I could reference it in future project sessions and provide context for tasks managed by Claude's agents.

I've heard someone say that AI is forcing engineers to document their work, and I completely agree.

The real bottleneck is clarity.

I've noticed that the more precise the documents I share with an agent are, the better the results tend to be. I spend less time going back and forth, correcting assumptions, and prompting it to try again. That saves me time and, ultimately, money.

I've started to appreciate just how valuable it is to have project documentation that I can reuse across sessions.

3. GitHub on my phone?!

Next, I was excited to download GitHub's mobile app for Android.

I'd already linked the v0 project to a GitHub repository, but I needed a system for managing and assigning tickets.

GitHub Issues was the obvious choice. It's free, accessible from mobile, and gives me a way to organize the work into smaller tasks.

I also wanted to experiment with assigning tickets to coding agents. GitHub has integrations for coding agents, including GitHub Copilot's coding agent, and I was interested in seeing how I could incorporate that type of workflow into my project.

In my case, I connected Claude Code to my project in GitHub. I would reference my tickets and assign them to agents within sessions. This was an alternative to creating a GitHub pro account.

I instructed the agent to stick to the scope of the issue, create a PR once the work was complete, and notify me when it was ready for review.

I also enabled notifications on my phone so I could keep track of the work.

And then it happened!

I checked the Github app and got to see that Claude Code had completed the task and pushed a PR to my repository.

I know this probably sounds like a small thing, but it was genuinely exciting to see the process come together. I had work happening in my repository without needing to sit in front of my computer the entire time.

4. My first ticket: auditing the homepage

The first issue I created was titled "Audit the church homepage against UI rules and list suggested fixes."

I intentionally started with an audit rather than asking the agent to make changes immediately.

The required output was a list of potential fixes that I could later translate into individual tickets.

I have a follow-up meeting with my client scheduled for later this month, and I thought having an initial assessment would help me move more quickly once I had their feedback.

I also wanted to establish a process for keeping the website accessible and improving its usability as the project evolves.

This is where the agent skills I'd been experimenting with came into play.

I'd created documentation outlining my rules for building well-designed components and maintaining consistency across the project. I then adapted a UI design skill from another source and referenced my own documentation so the agent would use those guidelines when completing the ticket.

I wanted to make sure its assessment was grounded in the design research I'd already done rather than having it come up with an entirely new set of rules.

The agent drafted its findings and saved the document in the project's /docs folder.

After reviewing the PR and the document it generated, I merged the changes and moved the ticket to Done.

One down!

5. The next ticket: creating a history page

The next ticket was a little simpler. I wanted to create a new page and route for the church's history, along with an archive of historical documents.

This is where having the previous documentation really started to pay off.

I had already established some standards for how new pages should be built, so I could reference exactly which files needed to be updated, which existing components could be reused, and what new components were needed to display the articles.

Before assigning the work, I actually took the ticket I'd written and asked Claude to help refine it.

I updated the description with the additional requirements and referenced the UI design skill I wanted the agent to use.

Starting a new Claude Code session to tackle the ticket was pretty straightforward after that.

When I got back to my desktop computer, I did have to resolve a merge conflict by hand.

I don't really trust the agent to handle those just yet, and I'm also pretty adamant about finding opportunities to build up my own mental map of the project's architecture and code patterns.

I want to retain enough muscle memory that if I need to jump in and apply fixes myself, I can.

Reviewing the code is a golden opportunity for me to do exactly that. I get to see how the agent approached the task, understand what changed, and test the UI myself.

Eventually, I'd like to add Playwright tests to the project as well.

After resolving the conflict and reviewing the changes, I merged the PR and we had a new history page!

6. Some thoughts on the process

Overall, I've really enjoyed the process of organizing sessions, collecting research, documenting changes, and having a little more bandwidth to schedule meetings and communicate with my client.

I think one of the biggest things I've learned is that having good documentation and breaking work into smaller tickets makes a huge difference when working with agents.

I'm still very much involved in the development process. I'm reviewing code, thinking about architecture, testing the UI, and making decisions about how the application should evolve.

But I'm also finding ways to delegate some of the work without losing sight of what's happening in the project.

And honestly, being able to check on a task from my phone while commuting is a pretty great bonus.

I'll talk more about how I'm using Claude agent skills in another post. I'm still experimenting with this workflow, and there's a lot more I'd like to explore!

On to the next item on the skill bucket list. 🍂