Pacman, Rebuilt
The scoped learning project where I figured out where AI-assisted development genuinely shines — and where a builder's judgment is still irreplaceable.
The story
When AI-assisted development got genuinely good, I didn't want to read about it — I wanted to feel where it shines and where it fails. So I picked a scoped, well-understood problem with hidden depth: Pacman.
It looks simple. It isn't. The original's ghosts each run different pursuit logic — one chases you, one ambushes ahead of you, one is unpredictable — and that interplay is what makes the game feel alive. Recreating it meant getting collision handling, tile-grid movement, and ghost state machines genuinely right.
What it taught me
- Directing AI like a senior engineer directs a team — clear specs in, working code out, review everything.
- Where AI accelerates (boilerplate, well-documented patterns) and where judgment is irreplaceable (game feel, edge cases, architecture).
- Confidence. Shipping this is what convinced me the multiplayer platform was achievable — one milestone at a time.
Outcome
A fully playable game — and a new build capability I've used on every project since. (The arrow keys still work.)
Build notes
Tech stack
- JavaScript
- HTML Canvas
- CSS
AI models used
- Claude Sonnet
- Claude Opus 4.8
- GPT 5.4–5.6