Start with the selector, because that is where the behaviour comes from. `.prose` compiles into rules that target the elements beneath it: `a`, `h2`, `ul`, `img`, and the rest of the tag vocabulary Markdown emits [6]. A hand-built CTA is, structurally, more tags in that subtree, and the plugin has no way to distinguish it from a link the Markdown pipeline produced [3]. The plugin is built for HTML you do not control [4]; embedding components you do control puts two styling authorities on the same elements [16]. Forty lines is a polite distance for a cause to sit from its effect [2].
The specificity story is the part worth testing. The modern plugin deliberately keeps specificity low using `:where()` [7], and a zero-specificity rule should lose to any class sitting on the element itself. Yet both reports above describe element-level utilities failing to take [1][10]. The writeup's account is that properties the component does not set inherit prose's values, and properties it does set are settled by CSS order [8]. Those two statements can only both hold if the utility declaration is emitted before the typography rules in your compiled stylesheet [1]. That is a fact about your build, not about the plugin, and it is the thing that decides whether this failure mode transfers to your repo at all. Check the order in the built CSS before you accept either the diagnosis or the ranking of fixes.
The framing that survives the check is ownership rather than victory [18]. Tier 1 is designing content-like embeds to cooperate: in this build the inline affiliate link is deliberately bold and underlined in the article accent [12], and the CTA settled on a light background, `text-stone-900`, and an accent border, all values prose has no opinion that conflicts with [13]. Tier 2 moves the rule to the container with element modifiers, `prose-a:text-emerald-700 prose-a:no-underline hover:prose-a:underline`, which the author calls the supported customization channel and the wrong tool for one special component [14].
Tier 3 is `not-prose`, and its cost is legible in the mechanism. Prose styles stop at the boundary [15]. That boundary cuts inward too: nothing inside the island gets the plugin's typography, so the island's owner declares headings, spacing, and link treatment by hand, including for any prose-like copy nested in a card [2]. For a product comparison card, an interactive demo, or a form, that is the correct trade, because those are not content [15]. For a component that is mostly words, you have swapped one context bug for a list of per-element declarations you now maintain.
