In the movie “50 First Dates”, Adam Sandler’s character falls for Drew Barrymore’s, whose rare form of amnesia wipes out her memory after every date. Working with Claude felt eerily similar. Each session, Claude gradually grasped the details of my code, understood exactly what we were building—and then promptly forgot everything the moment we closed the chat. Just like Sandler’s character, I learned a few tricks to help Claude “remember” our past interactions (more on that later).
Two weeks ago, I dabbled with “Vibe-coding”, and wrote about the experience here. In a day or two, I was able to put something together that was pretty cool. But it didn’t feel functional enough. The code was relatively simple, the app more so. The question remained: Could I build a true decoupled API-first application with AI generating most of the code?
I cleared my calendar—not entirely, I still had my day job—but for five to six hours each day, I dove deep into the world of AI-powered coding. To be clear, this wasn’t my first coding rodeo. Early in my career, I personally built most of the software for my first startup, a CMS company, and later wrote and spoke on coding and architecture. But after 15 years managing large engineering teams, budgets, and strategic roadmaps, my hands-on coding days had become rare.
My goal was straightforward: see if modern AI tools could bring me from an engineering executive back to a coder, enabling me to quickly build production-grade software without writing every line myself. We’re not yet at a place—and perhaps never will be—where deep understanding of code isn’t critical. But the crucial distinction is this: I didn’t actually write most of the code. Claude Code, GPT-4, and Augment Code handled that, guided closely by my experience, architectural decisions, and precise instructions.
Could the Hyperdev model genuinely accelerate software creation, or is it just another tech hype cycle?
I set myself a challenge to find out.
Spoiler: I shipped my first “production” build in six days. Over 40,000 lines of “production-ready” code. (See screenshots and diagram at the end of this article)
The Challenge: Can I Still Code?
My mission was clear: build a working, full-stack, AI-powered travel assistant from scratch—in one week. TripBot4, as I named it, was my fourth iteration, each taking what I’d learned from the previous versions, starting fresh with all new code, new architectures and evolving AI toolsets. At today’s blistering pace of AI innovation, experimentation isn’t optional—it saved me from several potential dead ends.
I strategically chose my stack:
AI Tools: Claude, GPT-4, Augment Code, OpenAI’s Assistants platform
Tech Stack: Next.js 15, React 19, Zustand (state management), Redis (persistence), Vercel (deployment), Apify for data scraping and Python for data processing. This stack seems particularly suited to AI/Vibe coding – the models seem to understand very well, and no surprise because by volume, it’s probably the largest example set out there. I haven’t tried Java, Ruby etc, but suspect the results wouldn’t be nearly as good.
IDE: I liked Webstorm, but VS Code was faster and had a great Augment plugin.
Data Infrastructure: Custom travel data aggregation pipeline, consolidating flight, hotel, and local activity data into unified, real-time streams.
I didn’t cut corners. TripBot4 is live, stable, and genuinely usable.
Myth: AI Can Build Your App for You
The marketing is seductive—“prompt and boom, your app’s done.” Reality isn’t that generous. Full-on “Vibe Coding” is a myth. It looks impressive—until reality punches you in the face. AI doesn’t magically build your app; it hands you raw materials that you must carefully shape yourself.
Here’s why pure prompting quickly breaks down:
Context Amnesia: Claude rapidly understood complex tasks—but forgot everything immediately upon session close. Each restart felt like another first date. There are lots of claims about ever increasing “Context Windows”, but in my experience you often need to lead AI by the nose and point it to the problem (and suggest a solution).
High-level Intent Misfires: AI often generated plausible yet incorrect or off-target solutions, especially as complexity increased.
Too Eager to Help: Without strict guidance, Claude frequently introduced unwanted complexity, requiring meticulous manual review.
My Solution: Think First, Prompt Later
The secret sauce wasn’t the AI itself—it was following a disciplined structure. My workflow became:
The Optimal Hyperdev Workflow
I discarded about a third of the AI-generated code. Claude and the others are fast enough that this waste is now just part of the workflow. I’m hearing lots of chuckles – “this is just good software practice!”. Yes, it is. The difference is that (a) I’m not doing the coding (mostly), and (b) AI speeds through the execution so quickly that you have time (and the necessity) to build the process armature around it.
Helping Claude “Remember”
Just like in the movie, where Sandler devised strategies to jog Barrymore’s memory, I learned tactics to help Claude “remember”:
Structured Context Summaries: Each session began with clear recaps: “Previously, we built X. Today, we extend it by Y.”
Architectural Signposts: Clear file structures and naming conventions reduced confusion.
Contextual Checkpoints: Frequent comments and explicit intent documentation quickly reoriented Claude each new session (each time I worked with Code, I would start a new ‘session’ - Claude only directly remembers work for that session. While it can make the sessions last, I found it more effective to work in discreet sessions designed to make a specific piece of code or solve a specific problem or bug)..
Tools & Workflow: What Worked Best
Each AI tool had distinct strengths:
Claude Code: Rapid, high-volume code generation and scaffolding.
Augment Code: Precision fixes and targeted refinements as well as strong IDE integration
GPT-4: Strategic research, planning and documentation partner. Was also helpful putting a second set of “eyes” on particularly tricky bits of code, or working with APIs that didn’t function as they should.
Just as important was IDE usage and robust version control (Git). “Vibe-coding” editors look attractive, but at AI-generated code velocity, granular control is mandatory. Quickly branching, merging, and rolling back saved me repeatedly.
High-performance deployment was effortless with React/Next.js (plus Shadcn+Tailwind) and Vercel. Without dedicated ops support, I achieved seamless deployment: clean builds, passing tests—boom, we shipped. Frictionless deployment was essential for rapid iteration.
Multiple IDE windows became common: debugging UI in one, building APIs in another. AI takes some time to do its thing, so I made use of that time by context-switching. Or taking calls, etc. I’d say out of a typical hour of “coding”, 50% of that time was waiting (with Claude – Augment isn’t doing as much and responds faster).
Rediscovering “Coding”
The energy generated from coding this way was surprisingly nostalgic. It vividly recalled my earliest developer days—when tools and frameworks were still immature, yet free from today’s crushing complexity. Hyperdev might actually return coding to its roots: problem-focused, rapidly iterative, and creatively invigorating.
In fact, my partner Joanie has never seen me as a developer—only as an executive. Watching me suddenly go heads-down, pulling late nights debugging and shipping code, genuinely unsettled her. At one point, she asked, “Who are you, and what have you done with my partner?”
Conclusions: Embrace Hyperdev or Get Left Behind
Here are the shocking takeaways:
Looking at the technical project description and AI’s analysis of what it would take a non-enabled engineer to develop, the gains are staggering — 10-20x. (That said, as I note below, my gut says it’s more closer to half of that.)
Hyperdev is real, transformative, and rapidly evolving—but no magical replacement for good engineering. It actively reinforces software fundamentals:
Well-architected, documented code
Clear separation of concerns
Comprehensive integration and unit testing
Clean data design
These principles aren’t optional; they’re essential for successful AI-driven coding.
Critically, you must embrace experimentation. Had I stuck with my initial vibe-coding approach, I’d be trapped in a technical dead-end. With AI’s rapid evolution, willingness to pivot isn’t just valuable—it’s essential.
Legacy stacks, outdated mindsets, and reluctance to adapt will kill your ROI faster than anything.
The Value of Throw-Away Code
I’ve always been a big believer in rebuilding. At my startup Citymaps, we made the hard decision to start again when our stack wasn’t serving our business needs. Hyperdev techniques on legacy code aren’t going to get you 5x productivity. 2x if you do it well, maybe. Which is something, but far short of the promise. But the productivity in a greenfield environment with modern stacks and tools is so great (5x or more), it may finally be the best argument to start again. At Citymaps, it took our team of six several months to train and rebuild. What if that could have been done by one or two people in two months? The financial equations are changing, and a ground up Hyperdev-optimized stack gives such productivity gains that the payoff will likely be within a short period of time. If your team is struggling under the weight of hard-to-maintain legacy systems, keep your business knowledge and data structures — throw away the code.
The Final Word
Start Hyperdev pilots today. Waiting isn’t caution—it’s surrender. Hyperdev won’t replace good engineers, but it will replace organizations who fail to adapt.
Like Sandler in 50 First Dates, navigating chronic forgetfulness isn't effortless—but once I learned the dance, productivity surged dramatically. It wasn’t a perfect romance, but I’d do it all over again because like Barrymore, Claude's and Augment's charm kept me coming back.
Acknowledgements
My colleague and friend Eddie Hudson of Divelement has opened my eyes to a new way of working with AI, stay tuned on that front!
Many thanks to my many colleagues and friends who helped proofread and suggest changes: Sean Graber, Amy Winter, Ophir Prusak, Andrea Kovac, Edan Golomb.
Special thanks to Joanie Dinowitz for her many suggestions and help with the final deep read to get this ready for publishing.
Appendix
Technical Project Overview
Project Timeline Summary:
Duration: 11 days (March 16-26, 2025)
Total commits: 97+
Code Volume:
Total files changed: 550+
Total insertions: 60,000+ lines
Total deletions: 16,000+ lines
Net new code: ~44,000 lines
Daily Productivity:
Day 1 (March 16): ~17,000 lines (initial setup, core chat functionality)
Day 2 (March 17): ~14,000 lines (UI enhancements, streaming functionality)
Day 3 (March 18): ~16,000 lines (state management, Redis integration)
Day 4 (March 19): ~4,500 lines (logging, version management)
Day 5 (March 20): ~3,500 lines (PDF export, icon compatibility, TypeScript fixes)
Day 6 (March 21): ~5,000 lines (chat UX improvements, itinerary API)
https://www.wsj.com/articles/the-ai-experience-is-going-from-50-first-dates-to-cheers-0dc6b9dd
Haha. Wonder where they got this idea from…