Nine If-Statements Beat One Model Call, Once Somebody Wrote The Cases Down

A developer rebuilding a small name tattoo tool reached for an LLM call to classify user input, then shipped a plain function with eight checks and a default branch instead, in a post published on dev.to ^[1][4][7]. The useful detail is the order of operations: he wrote down the cases he actually cared about before he wrote the prompt, and after that there was nothing left for a model to do ^[5][6].

The tool takes a line of text and shows a set of lettering styles, and at first every input got the same treatment ^[2]. The inputs are not the same shape: a single short first name, two initials with periods, two names joined by a plus sign, a name followed by a date, a hyphenated name, a two-word phrase ^[3]. Serving identical recommendations across that range started to feel wrong ^[2].

Because the product already used AI elsewhere, another model call was the path of least resistance ^[4]. What killed it was the case list. The author needed to distinguish initials, initials plus date, name plus date, two names, hyphenated names, multi-word text, and short, medium and long single names ^[6]. That is nine categories, not arbitrary human language ^[5][6]. The function that replaced the model call trims the string and falls through ordered guards: initials-with-date, name-with-date, initials, name pair, then a hyphen check, then a whitespace-split check, then length under six characters, then length under ten, then long ^[7]. Eight conditionals plus the fallback return produce exactly nine outcomes, one per enumerated case ^[16]. Note the ordering: the two compound date cases are tested before the bare initials and name-pair checks, so a compound input cannot be swallowed by the rule for its own components ^[17]. The medium band is four characters wide, from six to nine ^[18].

The scope discipline matters as much as the rules. The classifier does not decide which style is correct; it decides which directions appear first, and the remaining styles stay visible ^[9]. That is what keeps the branch count low, because a reordering function can be wrong without being harmful.

The stated payoff is unglamorous and checkable: no inference cost, no network latency, identical output for identical input, reproducible edge cases, a visible reason for every recommendation, and product changes that are code edits rather than prompt edits followed by hoping the model reads them the same way ^[10]. The author's conclusion is that the problem was not fuzzy enough to justify a model ^[11]. AI stays in the workflow where the request is genuinely open-ended, for custom compositions with flowers, symbols and flourishes, while the lightweight preview remains deterministic ^[12].

Two things to watch. The real implementation already carries extra checks around dates and separators beyond the simplified version shown ^[8], and that is where rule sets grow: every new separator, locale or date format is another guard, and the honest test is whether the branch count stays near nine. Second, the published recommendation map covers only four of the nine classes ^[15][19], so the ordering logic is less finished than the classifier. The transferable part is the screening question the author now applies before adding a model: not whether AI could handle the task, but whether the input space contains enough uncertainty to need it ^[13].