Most '100 best prompts' lists are useless. Here's the test, and the prompts that survive it.
I've sat through too many "100 best Claude prompts for marketers" listicles. They've all got the same shape. A wall of generic instructions, "act as a senior marketing strategist and write a tweet about my product", with no example output, no refuse conditions, no spec for what good looks like. You paste them in, you get back something that's roughly the model's default, you spend twenty minutes rewriting it, you wonder where the time went.
The good news: the test for whether a prompt is worth keeping in your stack is pretty short.
Every generic prompt you copy and run is tokens and time spent producing the model's default. You're not getting a brand voice. You're not getting your competitor's positioning. You're getting whatever Claude reaches for first, which is (by design) the average of the training data.
If the prompt doesn't ground the output in your context, you're paying $20/month for autocomplete. Don't.
1. It specifies the output shape, not just the role. "Act as a senior strategist" tells Claude nothing about what you want. "Output a markdown table with these five columns, then a 150-word executive summary, then three positioning hypotheses with kill criteria" tells Claude exactly what to produce. The first is decoration. The second is a contract.
2. It has refuse conditions. A good prompt tells Claude what to refuse. "Refuse to declare a trend from fewer than two data points." "Don't recommend comprehensive briefs, comprehensive is the loser's strategy." Without refuse conditions, Claude's default helpfulness gradient takes over. It will give you a confident answer to a question the data can't support, because that's what helpful looks like.
3. It grounds on real context. A prompt that asks Claude to "analyze my competitors" without you pasting actual competitor content produces inferences dressed up as findings. A prompt that requires you to paste the homepage hero, the pricing page, and a few G2 quotes produces real analysis. The difference is whether the prompt forces context or invites Claude to imagine it.
4. It marks its own uncertainty. Good prompts ask Claude to flag what's inferred vs. sourced. [INFERRED, verify against {{specific source}}] in the output is a feature, not a hedge. Marketers ship faster when they know which sentences need a fact-check and which are grounded.
Four of these are in the free library. The other six are in the paid packs, but the test for whether they're worth $29 is the same as above. Read the free four, run them, see if they feel different from generic listicle prompts. If they do, the paid ones will too.
Forces you to paste real competitor content. Outputs a structured matrix, a 150-word executive summary identifying where white space sits, and three positioning hypotheses with explicit kill criteria. Read the full prompt →
Three voice attributes, each one named with its opposite. Sentence rhythm patterns. Vocabulary lists (15 words to use, 15 to avoid). Six on-brand vs. off-brand pairs. The output is itself a system prompt you paste into a Claude Project so every future content prompt inherits your voice. Read the full prompt →
Turn the top 10 results plus a target query into a brief that beats them on something specific. Includes a "what NOT to write" section that's worth the brief on its own. The anti-padding clause ("don't recommend comprehensive briefs) comprehensive is the loser's strategy", is the line. Read the full prompt →
Refuses to declare a trend from fewer than two data points. Asks you to specify the decision the readout will inform. Distinguishes measurement artifacts from real changes. Read the full prompt →
The system prompt IS the campaign brief template. Claude refuses to output until every required section has real content. Interviews you back if your inputs are thin. Lives in the Campaign Planning pack.
One pillar article in, 6+ channel-native assets out, each one matching your voice (because you pinned the voice guide in the Project). Content Creation pack.
Distinguishes signal from noise. Refuses to call something a trend without weekly cadence data. Market Research pack.
Outputs a defensible content cluster with internal linking map, not a list of 50 generic keywords. SEO pack.
Names statistical significance honestly. Refuses to call a 4% lift on an N=200 test a winner. Data Analysis pack.
The shared-brand-context prompt the agents all read from. Produces a positioning statement with a kill criterion attached. Brand & Messaging pack.
Once a month, take every prompt you actually use and run it against these four questions:
Most prompts fail at least two of these. Cut them. The library is better when it's smaller and sharper.
If you want the audit done for you, that's what the paid packs are, every prompt has already passed this test. The 4 free ones above are the proof; the other 26 follow the same shape.
No email, no signup, no drip. Four full prompts with real sample outputs. Same shape as the paid 26, so if these earn their place in your stack, the rest will too.
See the 4 free prompts →