← THE MAPS P A C Y A P P S
EXPANSION · MAY 10, 2026
  EXPANSION  

Pure Vibe Programming Trials

MAY 10, 20264 MIN READESSAY

There's a mode of working with AI that gets called "vibe coding" — where you describe what you want and let the model write the code. I've been running a more structured version of that experiment, and I want to log it here as it happens.

My previous work with AI — writing, research, even parts of Secret Stuff — was a partnership. I brought the code, the domain knowledge, the taste. AI brought speed and breadth. Call it 50/50.

The question I'm now asking: can AI be the developer, while I stay in the seat of project lead and product owner?

Not just a faster way to type. A fundamentally different division of labour.

Trial 01 — Checkpoint

Checkpoint is a Safari extension that PIN-gates websites. Small scope, well-defined problem, no novel algorithms. A good first trial.

I wrote the spec. I described the UX. I made every product decision. Claude Code wrote the extension — manifest, content script, popup, PIN hashing, session locking — top to bottom.

It worked. Shipped. The code is clean. I barely touched it.

Verdict: at small scale, with a clear brief, AI as sole developer is not just viable — it's fast.

Active Trials · In Progress

Two experiments running in parallel. Different domains, same question.

WEB · NEXT.JS · GOING WELL
SpacyApps Site

· Design — Claude Design · Programming — Claude CLI · Content review — Grok · UI tweaks from live testing — fast and clean · Each tool stays in its lane

MAC · SWIFT · IN PROGRESS
Conduit

· More friction than the website · Structure and file organisation had to be directed · Refactoring required explicit prompts · Native domain is less forgiving · Ventured back into the code

The contrast is the finding. Web iteration and native app development are not the same experiment.

Design
Claude Design — direction set visually, not in code
Programming
100% Claude CLI — no manual coding
Content review
Grok used for review and refinement
UI iteration
Live testing → describe fix → done. Fast.
Developer mode
Never re-engaged — product seat held throughout
The gap between "AI can build it" and "AI knows when to clean it up" is where the human still lives.

What I'm Watching For

  • Where does AI get confident but wrong — and can I catch it from the product seat?
  • What decisions genuinely require code-level understanding vs product judgment?
  • Does the quality hold at larger scale, or does it degrade without a developer in the loop?
  • What does the final ratio actually look like — is it still 50/50, or something closer to 10/90?
The log continues. Go Nutz. — Walter Mak
fin —
“Go Nutz” — Walter Mak
MORE FROM EXPANSION
← back to the map
MAY 9, 2026 · 6 MIN

Mastering Human-AI Collaboration

AI adaptation isn't about resistance — it's about thoughtful partnership. A practical guide to staying ahead by keeping your own thinking in the driver's seat.

MAY 13, 2026 · 4 MIN

The Comments Next Door

The age-old debate about where to document your code — wiki, README, .md, commit message. The answer was always next to the code. AI just made it non-negotiable.