From hater to coder

18 months ago, I barely scraped a C in introductory college coding and swore off development entirely. As an Economics grad, my plan was strictly research.

Then I landed a Revenue Operations role.

Suddenly, my job demanded working automations: Clay workflows, HubSpot APIs, and dashboards that actually had to function. Opening VS Code for the first time felt completely foreign.

Instead of quitting, I used Claude as an infinitely patient technical pair-programmer. Fast-forward to today: I have shipped 20+ automations managing 5,800+ live contacts.

The traps

API drift: Models routinely invent endpoints or use deprecated syntax for fast-evolving platforms like HubSpot and Clay. Always verify against official docs.

Scale & rate limits: Scripts that run fine on 5 test rows often crash or trigger 429 errors when scaling to thousands of records. You must explicitly instruct the LLM to handle backoff delays and batch processing.

The prompt-loop trap: When code fails silently, prompting the LLM repeatedly often yields worse spaghetti code. Learning basic console logging is faster than relying entirely on AI to debug itself.

Question for makers:

If you use LLMs to build internal tools or automations, what guardrails or testing steps do you rely on to catch API rate limits and hallucinated endpoints before pushing to live data?