Refactoring English: Month 14

Highlights 🔗︎

• A new strategy for finding book readers is having positive results.
• I had a breakthrough experience by letting an AI agent run in unrestricted mode.
• I’ve been using AI to correct decisions I regret about my tech stack.

Goal grades 🔗︎

At the start of each month, I declare what I’d like to accomplish. Here’s how I did against those goals:

Publish three chapters of Refactoring English 🔗︎

• Result: Published two new chapters
• Grade: B-

I’m still having the same problem as last month where I’m getting too excited about AI experiments and letting it eat into my writing time.

Publish my 2025 annual review (year 8) 🔗︎

• Result: Published it
• Grade: A

It’s done! I’m happy with how it turned out. I’m now finished with a string of posts that I felt pressure to publish by a certain date, either because of tradition or time-sensitive subjects.

Refactoring English metrics 🔗︎

I published “The Most Popular Blogs of Hacker News in 2025” at the beginning of January, and it did far better than I expected. It gave January the book’s highest visitors and revenue since the Kickstarter last year.

Other revenue is down, as I haven’t had any editing clients, and the current owner of TinyPilot wrapped up their sponsorship of the book in 2025. I’m not planning to pursue any more professional sponsors because companies don’t seem that interested, and it’s easier to focus on readers.

Success by writing about other writers 🔗︎

“The Most Popular Blogs of Hacker News in 2025” is a continuation of a strategy I’ve been exploring for about six months. It boils down to: celebrate other software writers.

At a party last summer, I met a romance novelist who self-published. It was interesting hearing about her work because we had similar problems despite writing about drastically different topics.

I asked her how she found readers, and she said, “That’s the question!”

She told me that she had great results by starting a newsletter to review other romance novels, primarily by indie authors. It created a virtuous cycle where:

1. Readers discover and buy her book after reading her newsletter.
2. Other indie authors enjoy seeing reviews of their work, so they direct their own readers to her newsletter, which increases (1).
3. Her subscribers discover other interesting novels through her newsletter.

I love strategies where incentives align for everyone, so I started looking for ways to apply that to my book. I started writing about other bloggers and book authors that I enjoy and publishing those on my blog, and that’s worked well:

Discovering the power of AI sandboxes 🔗︎

I had a revelatory experience a year ago when I first started using an AI agent. At that point, I just let it edit files while I watched, and I thought that was a scary amount of control to give an LLM.

For the past year, I’ve mostly been coding with Cline, an AI agent extension in VS Code. It sped up a lot of my workflows, but I also micromanaged it aggressively because I didn’t trust it to perform arbitrary actions on my dev machine.

A few weeks ago, my friend okay showed me his AI workflow with Codex, OpenAI’s terminal-based AI agent. okay lets Codex edit files and run commands without direct supervision. It made me realize how much time I’ve been wasting babysitting my AI agents and dealing with hangs in Cline.

okay said that he sometimes lets Codex work unsupervised for an hour or more, and I couldn’t believe it. If I gave Cline a task that would take more than 10 minutes, it would either hang the UI, go down the wrong path, or explode my costs. But Codex is flat-fee rather than pay-per-token, which means you stop thinking about costs.

I still don’t trust any AI agent to run amok on my real computer, so I set up a custom sandbox for running AI agents on my machine. I go to the directory for one of my projects and run my custom command: sb. It spins up a rootless Podman container that has no access to my local network and can only see the current working directory. It has Codex and Claude Code pre-installed and authenticated with my accounts.

With AI in a sandbox, I was fine giving it full permissions to edit files, install applications, etc.

And wow, what a difference!

Seeing an AI agent run with full permissions was another breakthrough moment for me. Previously, if I said, “Make a bar chart of my income from the last 8 years,” about 30% of the time, the AI agent would implement something partially wrong. I’d have to check the result myself and say, “No, the bars are misaligned. Fix it.” But when the AI agent has root access in its own sandbox, it can spin up a test server, view the page in a browser, and iterate independently until it completes its task.

And then I heard about Ralph Loops. I still haven’t found a good explanation, so I’m not sure if I’m doing “official” Ralph loops, but here’s what mine looks like. I run a bash script called ralph-loop that contains this simple code:

#!/usr/bin/env bash

rm ALL-DONE.txt || true

while true; do
  cat AGENT-WORKFLOW.md | codex exec

  if [[ -f "ALL-DONE.txt" ]]; then
    echo "ALL-DONE.txt detected. Exiting."
    exit 0
  fi
done

And AGENT-WORKFLOW.md looks like this:

1. Pick the top task in TODO.md and begin work on it
   - If no actions remain, write a file called ALL-DONE.txt to the current
     directory, and exit.
1. Complete the task and delete the entry from TODO.md.
   - If the task is unachievable, explain why in the commit message.
1. Commit the changes with a detailed commit message explaining what you
   changed, why you changed it, and what impact it had.

And then I just create a TODO.md file with a list of tasks. Some of the tasks involve creating follow up tasks, so the list grows and shrinks with the agent’s progress.

The Ralph Loop has allowed me to run AI agents autonomously for 10+ hours unsupervised. It’s surreal to come back to my computer in the morning and see that the AI agent completed all the work I assigned it while I was sleeping.

AI is great at porting code 🔗︎

From experimenting with AI over the past few months, I’ve noticed it has the most impact when:

1. You can objectively define the success criteria.
   • e.g. “Find the cause of this crash” is objective and definable whereas “make this landing page better” is not.
2. The AI agent can verify success independently.
   • e.g., “Visit the page in a browser and verify the background turns blue when you push the button.”
3. The problem is such that a human with junior-level software knowledge could solve it with a search engine, experimentation, and patience.

Here are some classes of software tasks that meet these criteria:

• Refactoring code that has automated tests
• Porting code from one language/technology to another while preserving behavior
• Compiling a project from source, installing any necessary dependencies
• Fixing code to make a test pass

Recently, I’ve been using AI to port code. I have some codebases where I wish I’d made different choices about my tech stack, but it was always too time-consuming to rewrite everything. But with AI, swapping out pieces of my stack is inexpensive and fast.

I’ve successfully ported code in several of my projects:

• Converted the Zestful website from Vue/Nuxt2 to vanilla HTML with Hugo
  • I got a GitHub alert saying that I had some dumb vulnerability through a transitive Node.js library I’d never heard of. I thought, “I’d love to never see these alerts again.” So I had AI rewrite the site in Hugo and plain HTML/JS/CSS.
• Ported PicoShare’s CSS framework from Bulma to Bootstrap
  • When I created PicoShare, I wanted to try out Bulma as a CSS framework. It was fine, but I prefer Bootstrap, so I kept using it everywhere else and always had to switch gears when I worked on PicoShare.
• Converted LogPaste’s e2e tests from Cypress to Playwright
  • I wrote the e2e tests before I discovered Playwright, and now I’m so used to Playwright that it’s hard to go back to Cypress.
• Converted the fusion RSS reader from Svelte to vanilla HTML + Go templates
  • This is just a proof of concept, as fusion isn’t my project, but I’d like to fork it to use my preferred tech stack. AI did a good job converting all the Svelte code to vanilla HTML and Go templates, but I’d want to get more test infrastructure in place if I were to port this for real.
• Converted the MeshCore web app from Vue.js to Flutter
  • This actually worked poorly because it’s missing the “AI can verify the result” step. I thought Flutter would emit semantic HTML for its web app output, but it actually generates a wonky Flutter-centric HTML dialect. And the MeshCore app depends heavily on an external hardware device (the LoRa radio), so I had to be in the loop a lot.

Side projects 🔗︎

StreamPreserve 🔗︎

I’ve thought about what I’d do in a situation where I was witnessing something I wanted to record on my phone, but there’s a possibility of someone stealing my phone and deleting the footage or destroying my phone entirely.

So, I made StreamPreserve, a web app that quickly moves critical video to a remote, secure server.

Here’s how it works:

1. I open the StreamPreserve app and begin recording video.
2. The app streams a low-resolution video to the backend server, tuning the stream quality to the available bandwidth.
3. The app records high-resolution video to browser storage in discrete chunks.
4. With any spare bandwidth, the app uploads high-resolution video chunks to the server while continuing to stream.
5. When recording stops, the app saves the high-resolution video as a download on the local device.
6. When the app is open and no recording is in progress, it syncs all high-resolution footage from the device to the server.

So, the idea is if I record something and someone smashes my phone, my StreamPreserve server would still have a low-resolution copy of the video. And if someone seizes my phone to stop me from recording, the web app still uploads the high-resolution footage in the background.

I lost my enthusiasm for the idea when I realized there are a few flaws:

• It’s a bad solution for recording hours of footage.
• Implementing it as a web app adds complexity and the possibility of losing footage. It should be a native mobile app, but I dislike mobile development.
• It’s not a great task for AI because it depends on using the browser camera API, which is annoying to fake in an AI sandbox.

Wrap up 🔗︎

What got done? 🔗︎

• Published My Eighth Year as a Bootstrapped Founder
• Published two new chapters of Refactoring English
• Fixed a crash in the PoDoFo PDF reader that I discovered using my PDF fuzzing workflow

Lessons learned 🔗︎

• Celebrating other software writers helps me find new readers.
• AI agents are significantly more useful when they run in an environment where they have root access and can install applications and search the web.
• AI is good at solving problems when you can define success criteria objectively, and the agent has a way of verifying success and self-correcting through iteration.
• AI is great at porting code from one technology to another.

Goals for next month 🔗︎

• Publish two chapters of Refactoring English.
• Schedule a live event for Refactoring English readers.