The Rule at the Moment of the Decision
Written down, a rule looks authoritative. It has a heading, it is phrased imperatively, and somebody senior signed off on it. That is what it looks like in the document. What matters is what it looks like at the moment the decision gets made — and at that moment it is a small quantity of instruction sitting in the same room as a very large quantity of task material.
Call it what it is: a contest for attention, decided by weight and proximity. The rule says "do not lean on the English." The English is open, finished, correct, and one glance away. Every sentence of it is a ready-made skeleton for the sentence you are about to write. The rule is thirty tokens; the counter-argument is four thousand, and it is not arguing — it is simply present, which is stronger.
This Is Not a Machine Observation
It is tempting to file this under "limitations of language models," and that filing is wrong in a way that costs you the general lesson. Human translators have known this for a century, which is why professional practice separates the drafting pass from the revision pass and often separates the people. Copy editors read cold on purpose. Proofreaders are handed the text without the manuscript it came from. None of that exists because humans are unreliable; it exists because anything holding a source produces output shaped by that source, and the shape is invisible from the inside.
Attention-bearing systems, of whatever substrate, do not get to selectively not-attend to the most salient thing in front of them. That is what salience means.
Three Escalations That Do Not Work
Once you see a rule losing, there are three natural moves, and it is worth knowing in advance that all three are the same move wearing different clothes.
Make it louder. Bold it, capitalize it, add "CRITICAL." This raises the instruction's weight from negligible to slightly-less-negligible while leaving the material untouched. Three generations of the rule did exactly this.
Make it longer. Explain the reasoning, add examples of the failure, cite the audit. Now the instruction is nine hundred tokens instead of thirty — and it has become material too, competing with itself, while the source is still open.
Make it repeat. Put it at the top and the bottom and in the middle. This helps a little and fails the same way, because a reminder is only load-bearing at the instant it is read, and decisions are made continuously.
What none of the three does is change what is in the room.