3 posts in this category.
AIDA classifies support tickets with an LLM, so the classifier has to treat every ticket as a potential attack. The fence around the ticket text is real, but it isn't the guarantee — the guarantee is that a successful attack still can't do anything.
An AI agent blamed a failing Playwright test on Next.js response caching. I traced the code and blamed middleware. We were both wrong about the observation: the test's "anonymous" context was quietly authenticated. Two stacked bugs, two confident stories, and the re-run as the only reviewer that counted.
A model call is the slowest, priciest, least deterministic dependency in your whole system, so you ration it like an unreliable third-party API you call as rarely as you can. The practical skill is putting the model only where fuzzy understanding is genuinely required and doing the boring 80% with deterministic code you can test.