A curated library is a refused library. Six categories of prompt I cut on sight, and why your stack should refuse them too.
The shortest path to a bad prompt library is "yes to every prompt that sounds plausible." Most prompt libraries you can buy or download are exactly that: a yes pile. Every category got something. Every popular search query got a prompt aimed at it. The result is a library where half the prompts produce slop and the buyer can't tell which half.
The cleaner discipline is refusal. The Marketing Prompt Hub started with about sixty candidate prompts. Thirty made it into the library. Six categories got cut on principle. Not because they didn't work, but because they shouldn't exist in any working marketer's stack.
Generic prompts in your library mean you have to read every output to figure out which ones are usable. The time you save on the easy outputs gets eaten by the time you spend rewriting the bad ones.
A small library you trust beats a big library you have to babysit. Five sharp prompts in your Project beats fifty generic ones in a Notion doc.
The bait: scale your content. The reality: thirty LinkedIn posts in the style of a famous founder are thirty LinkedIn posts in the style of nobody, because the style of a famous founder isn't extractable from a few public posts. You get a costume, not a voice.
Voice transfer that works requires real samples of your writing, paired contrast (what's on-brand vs off-brand), and a refuse condition that blocks the model from inventing patterns. The Hub's voice extractor is built for that. The "write 30 in someone else's voice" prompt is built for the LinkedIn algorithm, which doesn't pay you.
"Act as a senior copywriter and write me a tagline." The role assignment is theatre. The model doesn't have access to a senior copywriter. It has access to its training data. What "senior" actually does in the prompt is shift the output toward longer, more confident, harder to disagree with. Confidence isn't quality.
Useful prompts give the model output specs and refuse conditions, not personas. "Output exactly five taglines, each under seven words, each with a paired anti-version that exposes the assumption underneath" produces a usable list. "Act as a senior copywriter" produces a confident essay you have to cut down by 80%.
The speed framing always sells. It also always produces more slop, faster. If your problem is volume (you're not writing enough campaigns, briefs, or posts), the bottleneck usually isn't your fingers. It's deciding what's worth writing. A 10x prompt accelerates the wrong half of the funnel.
The marketers I respect use Claude to slow down at decision points: more critique passes, more "what's the failure mode of this campaign," more pre-mortems. The 10x prompt skips that part, then charges you the slop tax later when the campaign underperforms.
"Comprehensive marketing strategist that researches, plans, writes, and analyses for any brief." A 4,000-word system prompt that promises to handle any marketing task is just five mediocre prompts mashed into one. You can't test it (which step failed?), you can't iterate it (changing one section affects the others), and you can't trust it (every output looks confident regardless of which mode it was using).
Better: small prompts that do one thing each, chained when you need a workflow. The Hub has four chained workflows for that. Each step is testable on its own, and a failure in one step doesn't poison the others.
This one's a brand killer. The prompt that writes "thoughtful" comments to leave on other people's posts produces the most identifiable AI bullshit on the platform. "Great point, [name]! I find that [restated point] is so true. What do you think about [related question]?" is an algorithmic giveaway, every time.
Even when the comment isn't visibly bot-like, the act of running an LLM over someone's post to generate a stranger-friendly response is the kind of thing that, when discovered, costs you trust faster than any campaign can build it. Don't.
This prompt produces costumes. It pulls quotable surface-level patterns (short sentences, contrarian framing, em-dash habits) and applies them to your content. The output reads as derivative, because it is.
Voice that compounds is your own voice extracted, sharpened, and made transferable to your team. The walkthrough for that is built around real samples and contrastive examples. The Steve Jobs voice prompt makes you sound like a knockoff. Knockoffs don't compound.
None of these prompts are technically broken. They run. They produce output. They look like marketing prompts. They'd inflate the Hub's count. They'd make for better SEO. "Get 100 marketing prompts" beats "get 30" on every keyword tool.
None of that is a reason to ship them.
The library you should keep is the one you can defend prompt by prompt. Six categories are missing from the Hub on purpose. The thirty that are in it earned their place by being run in real client work, audited against the five tests, and kept only when they passed.
Refusing to write a prompt is not a gap in a library. It's the feature.
Open whatever you're using to store prompts (Notion doc, Claude Project, Skills directory). Scan for any of the six categories above. If you find them: cut them.
If the deletion makes you nervous ("but what if I need it later"), that's the test. The fact that it might be useful for some future undefined task is exactly why it produces generic output today. Cut it. The library you trust compounds in value. The library you don't trust costs you the time it takes to verify every output.
Refuse more.
Each one passes the five-test audit. Each one was used in real client work before being added. Four are free with no email gate. Read them and tell me which categories you'd cut from your own stack.
See the 4 free prompts →