Almost any problem you bring to an AI has been worked on before. The most expensive way to solve it is from a blank page; the cheapest is to find how it was solved and check whether that fits you.
A model will happily design a system from scratch in a minute and will not say “this exists, it is called so-and-so” unless asked. So “has this been done?” is the second question, before any “how”: name the problem in its field’s words, find two or three existing solutions with real sources, then decide — adopt, adapt, reject or build. An existing solution is also a map of the pitfalls its authors already hit.
Repeating a solved problem on purpose — to learn, to reproduce, to check — is legitimate; only repetition out of ignorance is waste. In code this is Choose boring technology and the Rule of Three seen from the other side.
Where it comes from. George Pólya’s How to Solve It (1945) puts it among the first moves of problem solving: “Do you know a related problem?” — and, when one is found, “Here is a problem related to yours and solved before. Could you use it?” In engineering the same instinct is Dan McKinley’s “Choose Boring Technology” (75): known solutions come with known failure modes.
An engineering example. Asked for “a rate limiter for our API”, a model will often write a fresh counter-in-Redis design with its own edge cases. The field already has names for this: token bucket, leaky bucket, the generic cell rate algorithm, sliding-window counters — each with documented behaviour at bursts and at window edges, and implementations in the gateway you probably run. Naming the problem in its field’s words turns a design task into a selection task, and selection is easier to verify.
Limits. A found solution was built for its authors’ constraints — scale, consistency, team size — and fitting it to yours is real work (44). A dependency carries its own cost in supply chain and upgrades (35). And a model asked for prior art can invent plausible papers and libraries; the existence of a cited source is itself a claim to check (56).