You put “use British English spelling” in the system prompt. The first three paragraphs are fine. By paragraph nine there is a color, and by the end there is an organization. The instruction was not ignored; it was outvoted.
The characteristic pattern is not uniform failure. It is a document that starts correct and degrades — and the degradation is usually inconsistent within the document, so you get colour in one paragraph and color two paragraphs later, sometimes in the same sentence as behaviour. Long outputs are worse than short ones, and a long conversation is worse than a single call.
A second symptom is domain-specific: the spelling holds in ordinary prose and fails in technical contexts. Code comments, API field names, CSS properties and library names are American by convention (color
is a CSS property; serialize
is what the method is called), and text near them pulls the surrounding prose across.
Both patterns point at the same cause, and it is not that the model did not read the instruction.
Each token is sampled from a distribution conditioned on everything in the context. The system prompt is part of that context, but so are the two thousand tokens the model has generated since, and so is the enormous prior from training data in which American spelling outnumbers British by a wide margin in almost every technical domain.
At the start of a response the instruction is close by and there is little else in the context, so it dominates. As the response grows, the local statistics of the text being generated carry more weight relative to a single instruction several thousand tokens back. And the drift is self-reinforcing in exactly the way described in mid-answer code-switching: once one American spelling is in the context, the conditional probability of the next one rises.
The key insight for fixing it is that spelling is not a mode the model is in. There is no British-English state that gets set and then holds. Each word is an independent draw, influenced by context, and a single instruction cannot beat a strong prior across two thousand independent draws. That is why “ask more firmly”, “put it in capitals” and “repeat the instruction three times” all produce marginal improvements and none of them produces reliability.
The differences are systematic, which is what makes a deterministic fix possible. There are seven productive classes plus a list of one-offs.
The most common mistake in a spelling instruction is asserting that British English uses -ise and American uses -ize. It is not that simple, and getting it wrong makes your instruction incoherent.
Oxford University Press house style, used by the Oxford English Dictionary and by a number of British academic publishers, uses -ize in British English on etymological grounds — organize, realize, recognize. Most other British publishing, including most newspapers and the Cambridge house style, uses -ise. Both are correct British English. The OED documents its own -ize convention.
Two consequences. First, decide which convention you want and say so by name — “British English with -ise spellings” or “Oxford spelling” — rather than saying “British spelling” and hoping. Second, if you choose -ise, note that a set of verbs takes -ise in both conventions because the ending is not the Greek suffix: advertise, advise, comprise, compromise, despise, devise, exercise, improvise, revise, supervise, surmise, surprise, televise. A blind ize → ise
substitution is safe; the reverse is not, because it produces advertize and surprize.
Prompting reduces the rate; it does not eliminate it. The reliable approach is layered.
color
property, a serialize
method name and a quoted American source must not be “corrected”. This is the step people skip, and it breaks builds.The general principle generalises past spelling: any output constraint that can be checked mechanically should be checked mechanically rather than trusted to an instruction. That applies to variety selection in Portuguese and Spanish for exactly the same reason, and to the broader case of output that ignores a language instruction.
One thing worth knowing before you tune prompts: the baseline drift rate differs noticeably between model families, because their post-training data differs in how much non-American English it contains. If you can run the same document through two or three models behind one key and count the substitutions your post-pass has to make, that comparison is cheap and it tells you more than prompt iteration on a single model will.