How I build Pagecord Pagecord developer Olly released the first version of the blogging platform on 23 March 2024, three days after the first commit, and has since shifted most development gruntwork to Anthropic's Claude Code, which he adopted in May 2025 after finding the Opus model "very, very good all of a sudden." The workflow now centers on Claude investigations and implementations, with Fizzy cards created via the Fizzy CLI to track ideas, while the developer focuses on product definition, design, and code quality. The first version of Pagecord was released on 23 March 2024 https://blog.pagecord.com/hello-world-5f5e111 , only 3 days after the first commit https://github.com/lylo/pagecord/commit/3b085c357e3f69ec3e9ab5d04624c8012ed061b0 . It was pretty basic. I’m fairly sure I wrote most of this code without LLM assistance since it was early days for ChatGPT, but maybe I queried it for a few things – I don’t actually remember. I must have started using the ChatGPT chatbot to assist me that summer. There was a period in June and July where not a lot happened on the repository, then it became more active. I also did a talk at Friendly.rb https://youtu.be/ZsvJKG3uzpU that September where I mentioned building Pagecord using AI, so my view on what was possible and where this was going must have changed that summer. Around August 2024 I moved on from using ChatGPT in a browser to using VS Code and the GitHub Copilot integration to assist with my coding. The product evolved at a decent pace, but I was questioning what to do with it https://olly.world/much-ado-about-pagecord-188827ff towards the end of 2024. My enthusiasm waned and progress was a little slow from that point, then it started to heat up dramatically in May 2025. This was when I started using Claude Code https://claude.ai/code , which was fairly new, and the frontier model at the time Opus had got very, very good all of a sudden. From there on in, the pace of development of Pagecord has been pretty rapid and I’ve landed on a consistent workflow. I find it both enjoyable and highly productive. Over the past 18 months, more and more of my development gruntwork has been handed to Claude, my LLM of choice, with my time being spent on defining what I want the product to do and how it looks, using the product to write about progress and ideas, while also keeping a close eye on the quality of the code being produced, which is something I’m still interested in despite a growing movement against it https://youtu.be/vDjW dRyKXY?si=auLXidjmIIPtYBEy . It has been clear to me for months that the days of hand-typing code into an editor are no more and – crucially – I don’t think that’s a bad thing I don’t have a strict backlog of features for what I want Pagecord to do. I’m fortunate that I can dogfood the product https://en.wikipedia.org/wiki/Eating your own dog food , so I understand what works well, what doesn’t, and I get to explore what might be possible. As customer numbers grow, I get feedback from people as well, but I’ve noticed this is starting to plateau now that Pagecord can do most of what personal bloggers seem to want from a writing platform. I still track “things I want to do” in a todo list, but it’s fluid. I use Fizzy https://fizzy.do for this because it’s free, open source, and it’s built by 37signals I used to work there, so it’s built by my friends . Here’s a recent snapshot: Whenever I find a bug, or when I think of something new I want to explore, I’ll create a card in Fizzy. Sometimes I’ll do this by hand, but sometimes I’ll get Claude to do it. For example, if I’m in a Claude session and an idea for additional work comes up in the conversation, I’ll say “park that idea for now but create a card for it so I don’t forget”. Claude has a Fizzy skill which uses the Fizzy CLI https://github.com/basecamp/fizzy-cli to talk to it. New cards go into the Maybe? column in Fizzy. When I start work on something, it moves into the In Progress column until it’s live and I’ve blogged about it. That’s when I consider it “done done”, which is when the card is moved to Done. When I decide to work on a feature I will ask Claude to do an investigation. For a larger feature I used to enter plan mode, but increasingly, as Claude has improved, I just “raw dog” the implementation. It works surprisingly well, but sometimes you need a few iterations to hone it. A prompt for a feature might look like this: I'd like Pagecord to support a newsletter add-on for customers who have more than 1000 email subscribers. It will be $5/month or $50/year per 1,000 subscribers, counted across every blog on the account. No regional discounts for this. Complimentary accounts can't have the add-on; admins are unlimited. Never refuse a reader. Over the paid allowance no card, or a failed charge : email the blogger, pause newsletter sends after a grace period, keep sign-ups open etc. Checkout and pricing page must state the add-on is charged automatically as the list grows, so customers have agreed to it. Settings Audience shows the current block count and what it costs. That’s all you need – Claude will prompt you for details if it needs more context. Pagecord’s AGENTS.md https://github.com/lylo/pagecord/blob/main/AGENTS.md is fairly comprehensive and works well enough, but I should probably invest more time refining it myself. Over time, Claude also builds up a memory that helps it nail style, approach, and everything that is of concern to me. As of today I use Opus 5.5 on medium or high effort for most things. It’s really good, and it munches far fewer tokens. I switch to Fable for initial designs, and when I really want high confidence and deep thinking on a problem or if I’ve got a big Fable allowance left before the weekly reset . Claude works fast and it can produce a large implementation in fairly short order. It produces a PR on GitHub and that’s where I’ll scan the code, rather than looking at the diff on my laptop. I’m a software engineer by training, and I’m still passionate about Ruby code, so I’m curious. I can feel the need to do this waning quickly, though, I’m not gonna lie. Sometimes I’ll go back to Claude after looking at the code and question why it made a decision. I’ll challenge it to simplify things, and so on. I find the models need a bit of pushback, even Fable. Increasingly Claude will push back on the pushback – I’m getting used to wiping the egg off my face these days After the review of the code which is only a few minutes , I’ll start playing with the feature locally on my laptop. I’ll try every permutation I can think of myself, looking for where the UI breaks, where the flow doesn’t feel right, and thinking of areas where it could be better. I’ll then spend time going back and forth with Claude, asking it to improve certain areas, rethink approaches, come up with new design ideas for UI elements and so on. I’ll use Opus for this, most of the time. When things look good to me, I ask Claude to do an adversarial review: Perform an adversarial review on this branch It knows what to do. If I had more budget, I’d love to have a ChatGPT subscription as well so I could get Astra to do the adversarial code review because I think playing LLMs off against each other is a great idea if you can afford it. That said, Claude is clever and will boot another agent with no context for the review so you do get something of an independent review even with the same LLM. It always finds something to improve. It can sometimes take 3 or 4 adversarials, which get increasingly nitpicky. After 1 or 2 it’s probably fine to ship. When a feature is ready, I don’t necessarily ship it straight away. Sometimes a PR will live for a few weeks while I mull it over. Some things I’ve built never ship. I find building a feature regardless is useful practice – it allows me to properly visualise how it looks and feels alongside everything else. Shipping a feature or a bug fix means merging the code to main, then deploying the code to the server using Hatchbox https://hatchbox.io . Sometimes I’ll deploy a feature branch if I want to watch it for a day or two, and have the option of rolling back. Usually, though, I will just merge to main and deploy that. Occasionally I will use feature toggles if I want to run it just for me for a while, or have a selection of beta testers poke at it, but I don’t do this often. I almost always write an accompanying blog post for the Pagecord Updates https://blog.pagecord.com blog as well. I’ll normally do this before I ship the feature, so it’s ready to go when I deploy. I’ll also write a help guide, which I usually get Claude to do. Help guides are not meant to be creative writing like a blog post, so I don’t mind how it reads. I do often tweak what it produces, but frankly it’s pretty good at writing technical documentation and I’m comfortable with that. Blog posts are my writing of course, and I think that’s important. Claude does the shipping, except for the deployment. I just say: land branch-name “Land” is a term that’s come from GitButler https://gitbutler.com/ , which I’ve been using for the past couple of months. It’s great for when you have multiple branches on the go. Once the code is on main, I click the deploy button in Hatchbox and out it goes. Hatchbox does have an MCP server so Claude can actually deploy, but I still like going into Hatchbox and watching the deploy happen. Pagecord is the most bug-free product I’ve ever worked on. The odd cynic has suggested, disparagingly, that Pagecord is a “vibe coded” product simply because it’s built with LLM assistance. They are talking complete nonsense, of course. Pagecord is a highly robust product because it’s built with LLM assistance by a software engineer with over 30 years of commercial experience That said, of course there are some bugs that crop up. Weird edge cases, unexpected behaviour, race conditions, performance pitfalls. They all happen because the internet is wild. If there’s a 500 error in the app surprisingly rare , it will appear in Sentry https://sentry.io and I’ll get an email in my inbox, which alerts my phone. Claude is connected to Sentry and Gmail via MCP so I can just say “please investigate the Sentry error in my inbox and fix it”. It’s either a non-issue, or work gets done to address the problem. Claude knows what to do – create a fix on a branch and put up a PR in GitHub. I’ll take a look, then tell Claude to land the fix. If the bug is already on a Fizzy card, the prompt is even simpler: I use a MacBook Pro laptop but a year or so ago I bought a mini PC for hacking around with Linux. Occasionally I’ll use this PC for development, but I mainly use the Mac. The PC is running 24/7, locked down via Tailscale, and I have a Claude session running on it which is accessible via remote control https://code.claude.com/docs/en/remote-control . This allows me to access Claude and the PC on my phone or iPad. This has opened up a whole new world since it no longer ties me to a laptop for doing work, specifically support. Previously, if I needed to do some support I would need to get the laptop to investigate. Now I can just ask Claude to do this from my phone. The phone is talking to Claude Code directly on the PC my “claude server” , so it’s essentially the same as being on the computer, but I can now do it from anywhere Running a SaaS product as a solo founder means being on call 24/7/365, so being able to use your phone to ask your agent to investigate and solve an issue is, honestly, liberating. Now that I have a Claude server running 24/7, I’ve been getting it to do more and more gruntwork for me. The most valuable task is my “Ops Note”. This is a daily note written overnight by Claude that tells me over breakfast who signed up, who left, how the app is performing, what bot attacks happened, what the server status is, and whether anything needs my attention. I love it A systemd timer runs a shell script at 2am that pulls down the previous day’s production logs, runs rake tasks https://github.com/lylo/pagecord/blob/main/lib/tasks/logs.rake to wrestle them into the report, which is then stored in Obsidian and emailed to me. This has been a game changer, and it has found many performance issues quickly fixed , bot attacks and other things I’d never have known about otherwise. The nightly job will create Fizzy cards for anything that needs attention. This could be and probably will be soon automated so these issues are fixed on a PR or even deployed if it’s an emergency , but for now I review them in the morning and use the old “fix fizzy card 453” prompt when I get a chance. It’s useful work for me to do. The fact that all of this can be managed just with my phone has changed everything, and I’m now focused on even more automation to make the platform as self-healing as possible with appropriate guardrails in place of course . This is the way of the future, I think. LLMs, Claude specifically, have allowed me to create a productive, enjoyable workflow to build and maintain Pagecord. They have freed me to focus more and more time on what really matters – improving the user experience, offering exemplary customer support, and honing the marketing, positioning, and brand. Today Pagecord makes a paltry $240/mo pre-tax profit, so it’s not worth me trying to spend 40 hours a week churning out features. If I can spend just 8 hours over an entire week, when it suits my schedule, then it keeps the whole project fresh and it doesn’t ever feel like a chore, a time sink, or a grind. Say what you like about LLMs, but they’ve been a godsend for solo developers like me trying to bootstrap a small business on the side. I plan to keep running Pagecord for years to come, and with this AI assistance and automation I think that’s now possible without it becoming a burden, even if the product only makes a few hundred dollars in profit a month. So many indie products have died an early death over the years because they were too much effort for too little return. It no longer has to be this way. Perhaps one day I’ll discover the secret sauce that brings thousands of paying customers to Pagecord, along with a decent annual income, but it doesn’t really matter if not. I can now sleep relatively easy knowing I have a product that people love and use every day, that’s reliable and resilient, and that can be operated by one person with a minimal amount of oversight. That feels pretty remarkable to me.