Field notes
9 articles
We Lost the Team, So We Rebuilt the PlatformThe story of how one of our members, Don, went from using Copilot as a fancy autocomplete to leaning on AI to learn Terraform overnight when our DevOps guy walked out the door, to building the admin screens my frontend team never had time for. Then a funding glitch took the whole team, and the three of us left made the bold call to rebuild v3 with Claude.
Well Hello Fable!Fable landed on my max plan for good and I'm thrilled, because I've been having a blast with it and making real progress. It's a wonderful cofounder for the thinking and reasoning work, less prone to wandering off, though I still hand the actual coding to Sonnet 5. I also tease the Obsidian vault that's slowly morphing into Virtual Rob as a Service.
Big Codebases, New ChallengesA 2,300-line PR landed from a sprint that took about 30 minutes, a first-pass Stripe webhook receiver with real test coverage and code I'm proud of, and I couldn't stop thinking about who is supposed to review a thing like that. This is the bottleneck we're all about to hit, especially with the giant enterprise refactors coming our way, and I've got an idea I'm going to build in the open this week.
Throw Fable at your notes while you canIt was Fable's last day, so I made the case for pointing it at the work it's actually great at, the connect-the-dots reasoning tasks rather than code. I dumped every note, post, course, and book I own into a single Obsidian vault and let Fable help me turn the mess into a shared memory space, while keeping the parts of my process that need to stay in my own head.
The Human's Scarce AttentionA cohort session got us onto the thing nobody has really solved yet: how do you review a multi-thousand-line AI PR without becoming the bottleneck that grinds everything to a halt? We chewed on smaller PRs, trunk-based development, and letting the tests stand guard, but the bit that stuck with me is that a reviewer staring at a huge diff isn't refusing to read code, they're refusing to reconstruct your intent from scratch.
Is it possible to squeeze good code from a cheap model?I wanted to know if a cheap, dopey little model like Haiku could write genuinely good code as long as I wrapped it in enough skills and instructions. So I gave it a 57-line CLAUDE.md, one prompt, and told it to build me a store while I went to the beach. Opus reviewed the result and found a mess of singletons and security holes, so I got stubborn and tried to make it work anyway.
Why Claude Writes Better Code Than MeA new YouTube video on the extra mile it takes to get high-quality code out of Claude, and the set of skills I built around exactly what I like to see. Members can grab the skills and put them to work in their own codebases.
Claude has some weird opinionsI noticed Claude Code had quietly stopped writing comments in my code, and when I dug in it told me my instructions contradicted the "coding harness". That sent me back down the old rabbit hole about whether comments are a crutch, a bit of dogma I've never bought, and I'm genuinely curious where you land on it.
Creating a Masterpiece with Claude CodeI decided to see how far I could launch the worst interview question in the world, fibonacci, with Claude riding shotgun. There was no script this time, just me and a dumb prompt, and I loved how it came out. Gemini reviewed the result and called it a masterclass, which still makes me grin.







