Generic is not a Claude problem. It is a prompt problem. The five patterns that produce vague answers, with the exact rewrite for each, plus two tools that catch them before you ship.
The pattern goes like this. You open Claude. You type something reasonable. You get back something that reads like the average answer to the average version of your question. You think the model is shallow. You try a different prompt. Same result. You consider going back to writing it by hand.
Claude is not shallow. It is producing exactly the level of specificity your prompt asked for. The reason the output sounds generic is that your prompt is generic. Specificity in, specificity out. The five patterns below are the most common ways marketers strip the specificity out of their own prompts without realising it. Each one has a fix.
Before you hit send, read your prompt and ask: "What would a reader who has read 100 of these instantly know to skip?"
If your answer is "nothing, this is too vague to skip anything," your prompt has no specifics yet. Claude will give you back the average. The fix is not in Claude. It is in the input.
Claude mirrors the structure of what you put in. Adjectives are the laziest form of instruction because they delegate the decision about what "engaging" or "compelling" actually means to the model. Claude defaults to the average meaning, and the average is the generic version.
A prompt that works without any inputs is a prompt that works without you. It is going to return the average answer for the average company in your category. If your prompt is missing brand voice, ICP, last quarter's data, or your actual situation, Claude is guessing at all of those. The guess is the generic part.
Without a refuse-condition, Claude's default is to be maximally helpful. For a question the inputs cannot support, "maximally helpful" looks like a confident, generic answer that fills the gap with averages. The refuse-condition reroutes that energy into telling you what is missing.
"Act as a senior marketer" returns an answer pieced together from the average opinions of the average senior marketer who has ever been quoted online. Which is, predictably, the generic version. The pattern is in one of the categories of prompts I cut on sight. The fix is not a more senior persona. It is removing the persona and replacing it with the actual decision you want made.
This is the lever covered in "Claude mirrors you". Terse, command-shaped, lowercase, no-punctuation input produces terse, command-shaped, generic output. Claude does what your structure suggests it is being asked to do, and your structure said "give me a quick generic answer."
If your output is generic, work through the five patterns in this order: source material, refuse-condition, persona removal, adjective removal, conversational register. The first two are the biggest levers. If you only have time for two fixes, do those two and most of the genericness disappears.
A prompt that passes all five is what the rest of the work in the library looks like. Each prompt in the paid library is built on these (plus output spec and uncertainty flagging, the two structural levers covered in the levers post) so that the output you get is meeting-ready, not draft-ready.
The prompt audit tool scores your prompt against the six structural levers in your browser, so you can catch the missing inputs and refuse-conditions before you send. The Slop Sniffer catches the verbal side: the AI tell-words, the hedging filler, and the horoscope phrases that signal "no decision was made here." Run a generic-feeling prompt through both. One of them will tell you why.
Open the audit tool Open the Slop Sniffer