This fall I decided to fill some time building a SaaS product. I had a brief lull in consulting work, and I wanted to level up using the latest AI assisted IDEs, experiment with Claude code and keep my skills fresh. Over the break I decided to shut it down (I ran out of GCP credits for the database instance). Here is a scatterbrained account of my journey and what’s next.
My vibe coding journey had started early — back in early 2024 a friend of mine at Groq told me to try out v0 from vercel. This was the first developer tool that actually sparked something in me. I was vibe coding geometry dash clones with my kid, and prototypes for work. My sleep was slightly affected and I was already was waking up early to tinker with harder problems and specs. Could I create a grafana clone? (yes). How about visualize vector embeddings in 3d space (sorta). So many TCO calculators…
In May I focused my attention to a few consulting projects and leveled up to doing some real engineering in Cursor. I hadn’t coded in 10 years, but combining my product experience with AI leverage for my clients felt right. I had a glimpse at a potential future path, but my brilliant wife told me to slow down. The summer had higher priorities — a month off in Europe.
Of course returning from a prolonged vacation left me refreshed, energized and excited to jump back in. But doing what? I went through a series of high profile interviews at the labs and tech companies…I was close and the rejection was painful.
I needed an outlet, and I wanted to improve my workflows. I really like CLIs so I wanted to try out Claude Code, but I needed to focus on a problem. This was the genesis of vibe check, and my first attempt to take on something complex and technical using ai coding assistants. I’d build a programming language (well a DSL at least). My journey had started.
I wanted to make this a post about what I learned, but honestly it’s so many things I can’t even (book a time if your curious). A word of caution to others looking to take this journey it was this point that my sleep began to suffer — but I was having so much fun! I wasn’t sleeping because I new my limit had reset and had to get back to it. I had tinkered with Claude Code, but really I was maxing out on Cursor at this point. I had reached Level 4.
I was following RPI, and moving at what I thought was a breakneak (I was actually only at the start of my exponential, the second row is where it really gets crazy). I leveraged my fledging developer preview to network, sneek into re:Invent, and learn learn learn, but it didn’t achieve a business outcome (yet, who knows).
So I decided to go back to my consulting clients and see if I could help them out. I am fortunate enough to have a network of applied technologists across fields and sure enough my client wasn’t happy with their search system. I wrote a proposal for a lift and shift, and I provided an initial price estimate of the new system. It should be dramatically cheaper but the point of the POC was to prove it.
Other claims about capabilities were made, and I got a little nervous this was going to be harder than I thought. I spent two weeks doing nothing but system design (smart decision in retrospect especially since Opus 4.5 was about to drop).
My agent prep was working. After the design phase I was absolutely flying. I worked up to the Thanksgiving break maxing out Cursor and updated to higher subscription. I had separate microservices going and my seperation of concerns was allowing for parallel streams. I was starting to get annoyed with Cursor and when Opus dropped over the break I knew it was time to switch out. I had reached Level 5 and was rapidly headed toward Level 6.
I delivered an on time the search system, that searched 10x the top-K of their current and it was roughly 1/10th the price. I was no longer just playing with fire, now I had a combustion engine. (I was also playing with fire and made a $4k operational mistake but not going to go into that on here).
Over the break, I focused on improving my workflows. Beads, ralph loops, worktrees, slack hooks, tmux all of it. Level 6 and 7 got unlocked and sleep became elusive. Building a programming language, a database. All of it seems possible now but one lesson I learned on vibe check is a product without a team can’t win.
So I am pivoting my time to helping teams adopt AI and embrace FAAFO (fast, ambitious, autonmous, fun and optionality). I’ll be continuing to chronicle it here.
Thanks for reading.
