And the five tests to apply to every prompt in your stack, including the ones you wrote yourself.
Open any "best AI prompts for marketers" listicle and you'll find some version of: "Act as a senior marketer and write a tweet about my new product." Multiplied by 100. Some of them are titled "The only ChatGPT prompts you'll ever need." Spoiler: you'll need more than that, because none of these prompts will produce anything other than the model's default.
I'm not going to be precious. Most of these lists are content marketing for prompt libraries, written by people who haven't run the prompts they're recommending. The test for whether a prompt is worth keeping isn't long, it's just rarely applied.
Every generic prompt you copy and run is tokens spent producing the model's default. You wrote a tweet badly, faster. You wrote a blog intro that sounds like every other blog intro, faster. The "speed" gain is offset by the rewrite time.
A small library of prompts you've actually tested beats a list of 100 prompts you've copied off the internet. Five great prompts in your Project beats fifty generic ones in a Notion doc.
Apply each one to every prompt in your stack, including the ones you wrote yourself. Most prompts fail at least two. Cut those.
"Act as a senior strategist and analyse this market" doesn't specify anything. Claude has to guess at what shape the output should take, and it'll guess at the average, which is a 2,000-word generic essay.
Versus: "Output a markdown table with these five columns: competitor / positioning / ICP / pricing / weakness. Then a 150-word executive summary identifying where white space exists. Then three positioning hypotheses, each with a kill criterion."
The second prompt produces useful output the first time. The first prompt produces something you'll spend twenty minutes reshaping.
Refuse conditions are where good prompts earn their keep. They clamp Claude's helpfulness gradient: the tendency to produce a confident answer to a question that doesn't have one.
Examples of refuse conditions that change everything:
Without a refuse condition, Claude defaults to maximally helpful: which, on a question the data can't support, looks like confident bullshit.
If the prompt works without you adding source material, the output will be generic. Period.
A "competitive analysis" prompt that doesn't require you to paste actual competitor content produces inferences dressed up as findings. A "voice transfer" prompt that doesn't require you to paste actual brand writing produces a voice guide that looks like everyone else's. A "performance readout" prompt that doesn't require you to paste actual data produces commentary that could fit any quarter.
If the prompt feels like it works with no inputs, that's the signal it's not going to give you anything useful.
Good prompts ask Claude to mark inferences vs. sourced claims. The [INFERRED, verify against {{specific source}}] tag in the output is the difference between a brief you can ship and a brief that has a wrong premise hidden inside it.
If every claim in the output comes out at the same confidence level, you'll either over-trust the output (and ship something wrong) or over-edit it (and waste time). Honest uncertainty is the third option.
The best prompts include a "skip this" section. The SERP brief tells you when the SERP is unwinnable. The performance readout tells you when small-N invalidates the conclusions. The competitive teardown tells you not to recommend a positioning you can't credibly take.
"What NOT to write" / "what NOT to recommend" / "when this analysis breaks down" are the sections that separate prompts written by people who use them from prompts written by people who think they'll be impressive.
Open your prompt library (Notion doc, Claude Project, wherever). Pick one prompt at random. Run each test against it.
If the prompt fails two or more tests, cut it. If it fails one, fix that one and keep it. If it passes all five, keep it without changes.
Do this once a month and your prompt library compounds in quality. It's the same audit a senior copywriter does on a junior copywriter's portfolio. Apply it to your prompts.
That's literally the value proposition of the paid prompt library. Every prompt has already passed the five tests. The four free prompts are the proof, read them, run them, see if they feel different from the listicle prompts. They do. The other 26 follow the same shape.
You don't need to buy ours. But you do need to audit your own. Generic prompts compound generic outputs. The library you should keep is the small, sharp one, not the big, listicle-shaped one.
Read them and run them. If they feel different from the prompts you're using now, the paid 26 follow the same shape. No email gate, no signup, CC BY 4.0.
See the 4 free prompts →