After discovering the protocol mismatch, I built a small Node.js adapter to translate between browser‑friendly WebSockets and device‑friendly Protobufs/TCP. AI helped me build it quickly.
But as the project grew, a new pattern emerged:
AI has limited short‑term memory.
Every time I asked it to refactor or extend the adapter, it would forget something important. It would miss an edge case. It would overlook a callback or a timing safeguard. Not because it was careless, but because it simply couldn’t hold the entire code base in its head.
So, I adapted my workflow:
- I provided the current source code before every change.
- I restated the invariant states and assumptions.
- I validated every update against real behavior.
AI became a fast, capable junior developer—one who needs frequent context resets.
The productivity gains were still there, but they depended on disciplined oversight.
What strategies have helped you keep AI “on track” in multi‑step work?
Leave a Reply