A frontend developer with eight years of experience at a small company, where there is no dedicated frontend team; he handles all three platforms, build processes, deployments, and midnight alerts single-handedly. Over the past six months, he has delegated every task possible to AI, sharing a practical observation: what truly makes me uncomfortable is not its ability to write code, but that the benchmark has been changed.
This observation is more specific than the common anxiety about replacement. Previously, he delivered one page per week, which the boss considered normal; now, colleagues using AI produce three proposals in a day, and the boss asks him why he is so slow. AI did not announce anyone's exit; it simply raised everyone's expectations by one level and left each person to chase that increment. The most difficult aspect for small companies lies in this: having to make judgments AI cannot perform while also patching up the parts AI gets wrong, leaving not a single hour of the day fully dedicated to thinking.
He drew a clear line. Tasks where AI is indeed effective share a common trait: they have standard answers. Regular expressions and various legacy compatibility codes, which previously took half an hour to research and ten attempts to test, can now be verified and used in two minutes; for a ten-year-old table page requiring Excel export last quarter, AI wrote the entire xlsx segment directly, and after adjusting two style points, it went live; pasting compressed error stacks allows for rough localization, saving half the search time; writing comments, documentation, and technical solution explanations for the boss are where the most time is saved.
However, what it cannot handle happens to be exactly what he deals with daily. Why this field cannot be deleted—three years ago, a client's data depended on it. This information is not in the code, but in human memory. Those 'ugly' writing styles in old code were intentionally designed that way to bypass a pitfall back then; if AI enthusiastically helps optimize them, the production environment crashes. AI does not sign off on deployment; when the boss says 'make it look more advanced,' AI is even more confused than he is.
His response consists of three principles. First, treat AI as an intern, not an outsourced contractor; read every line it writes, and after reading, ask yourself: could I write this segment myself right now? Second, step forward; do not just translate requirements, but start engaging with business logic, calculating costs, and proposing solutions, because AI cannot answer whether a requirement should be implemented at all. Third, reserve a territory only you possess; organize the pitfalls of this legacy project into documentation, which is something AI will never obtain.
Regarding task delegation, he currently firmly refuses to hand over four categories to AI: payments, permissions, data deletion, and online hotfixes, while being flexible with everything else. He did not provide explicit reasons for this classification, instead asking readers at the end of the post if there is a more reasonable way to divide these lines.