On March 10, 2024, I opened an issue in one of my repos called "Clean up Makefile". Then I left a comment on it:

Two minutes later there was a pull request with the clean Makefile.

Today, this is common practice for most developers. Back then, my mind was blown.

That was the very first pull request opened by PR Pilot, the AI coding agent I'd been building on the side. Eleven months later, Anthropic released Claude Code. This is the story of what I built, why I'm still proud of it, and why nobody used it.

In 2023, building with LLMs was still very clunky and difficult. GitHub Copilot was the first thing that actually enabled development with AI, but it was more like a smart autocomplete. Everything else meant copying code back and forth between ChatGPT, GitHub and whatever other system I was working in. I was constantly context switching.

So I built my own version of what people nowadays call an agent: you give it a request, and the agent takes care of it. It also had to live where I already worked, so it needed to integrate with GitHub and all the other tools I was using every day.

The first version was a GitHub App. You wrote /pilot followed by a request in any issue or pull request comment, and it would answer questions, write code or search the internet for information. Every task ran in its own throwaway container in the cloud, and every reply came with a link to the log and an Undo button (just in case it got a little too creative).

PR Pilot answering a /pilot comment. Notice the Log and Undo links at the bottom.

I posted it on Hacker News on March 14, 2024. It got 7 points. Devin had been announced two days earlier. Timing is everything.

Next, I built a command line tool called pilot. Version 1.0.0 shipped on May 27, 2024. By late September it had seen 70 releases, 39 of them in June alone. In August I added skills: reusable instructions you could teach it for your own workflows. Sound familiar? :)

This is where it got interesting. I realized that the system I was building was actually making me more powerful and more efficient. It let me do higher-level thinking instead of getting caught up in the details of writing and deploying code.

One of my proudest moments was using the CLI in a very similar way to how I use Claude Code today. Here is a session from September 2024, going from an error in Sentry to a GitHub issue, a Slack message, a fix and a pull request:

Me and pilot in my terminal, September 2024. Five months before Claude Code.

The CLI felt very native and worked with any stack I threw at it. I had built myself a power tool that just wasn't available out there.

On February 24, 2025, Anthropic released Claude Code as a research preview. In May it became generally available, and you could tag it on GitHub pull requests "to respond to reviewer feedback, fix CI errors, or modify code."

My first honest reaction:

It was a moment of pride. It was also a moment of deep resignation.

I wasn't the first. Aider was already working in the terminal in May 2023. Sweep was turning GitHub issues into pull requests by July 2023. Devin beat my Show HN by two days, and GitHub previewed Copilot Workspace about six weeks later.

PR Pilot also had real limits. It couldn't run a shell or its own tests (tests ran in CI after the fact). And GPT-4, the model doing the work, had what my own docs called "the baked-in tendency to leave placeholders." The more complex the task, the sloppier it got.

Models back then weren’t trained on agentic workloads. This was before “agentic” was even a term.

Still, I used it every day, and for my own work it did the job.

Here are the honest numbers. PR Pilot opened 282 pull requests, nearly all of them on my own repos. I posted it (and the tools I built on top of it) on Hacker News ten times in 2024. The best post got 26 points. I made it free for open source projects, then made the whole thing open source. I priced it in credits instead of a subscription. I rewrote the homepage around personas and finally renamed it to Arcane Engine.

None of it moved the needle.

Looking back, I just did not know how to tie it to actual problems that people had. I didn't know anything about marketing, how to really present a thing, or who to even present it to. Selling another tool to developers felt like an uphill battle, because that's what everyone was doing at the time. Developers (myself included) were being bombarded with new AI solutions, and my messaging drowned in that.

I built a cool tool. I didn't build something that increases your ROI by X.

The repos are archived now, but they're still public. You can browse the CLI and the docs if you're curious what an AI agent looked like in early 2024.

These days Claude Code is part of my daily work, and every now and then I catch myself typing a request into my terminal and thinking: I've seen this before 😄

If you're building something ambitious and want someone who has shipped (and learned the hard way what it takes to get people to care), reach out and let's connect.