Most prompt libraries are junk. They’re either written by people who don’t do marketing, or they’re written by people who do marketing but stopped shipping six months ago. The result is the same: polite, generic scaffolding that gives you polite, generic output.
The prompts below are the ones I actually paste in every week, across Claude and ChatGPT. Some work better in one than the other. I noted which. None of them are clever. They’re all boring in the way that good operator tools are boring — they save a morning and compound over time.
You can copy-paste each one straight into the tool, swap the bracketed placeholders, and ship. These are designed for a lean marketing team, but most work for a founder doing their own marketing too. If you want context on why the two-human AI-marketing setup works, I wrote about that in the 2-person stack piece.
1. The audience-research drill
For understanding who you’re writing for before you write anything. Claude handles this better than ChatGPT because it sits with ambiguity longer instead of jumping to conclusions.
Act as a sharp, skeptical research analyst. I'm going to give you a product description and three real customer interview transcripts. Your job:1. Tell me the three strongest themes across the interviews — the things customers actually care about, in their own words where possible.2. Tell me the one thing I'm probably wrong about based on what they said vs. what the product description claims.3. List three questions I should ask in the next round of interviews to pressure-test my assumptions.Product description: [PASTE]Interview 1: [PASTE]Interview 2: [PASTE]Interview 3: [PASTE]Be blunt. If the evidence is thin, say so.
The “tell me the one thing I’m probably wrong about” line is the one that earns its keep. Without it, Claude will flatter your positioning back to you. With it, you find out that your “premium” segment is actually price-shopping and you need to rework the hero page. Anthropic’s own guide to prompt engineering is the best plain-English primer if you want to go deeper.
2. The content repurpose chain
For turning one long-form piece into five short-form variations. ChatGPT runs this one faster than Claude and doesn’t require as much guardrail-setting.
I'm going to paste a blog post below. Your job is to spin it into:- One LinkedIn post (120-200 words, first person, ends with a question that invites replies)- One Twitter/X thread of 6-8 posts, each under 270 characters- One short-form video script, 45-60 seconds, talking-head format- One email subject line + preview text combo, for a weekly newsletter- One pull quote (the single most interesting sentence, rewritten to stand alone)Match the voice of the blog post. Don't use hashtags in the LinkedIn version. Don't use emojis anywhere. Don't start any of them with "Ever wonder" or "Here's the thing."Blog post: [PASTE]
The “don’t start with ‘Ever wonder’” line is a humanizer hack. Every AI model has a small set of opener templates it falls back on. Explicitly blocking the worst ones at the prompt level cuts the cleanup time in half.
3. The competitor teardown brief
For weekly competitor scans. Claude is better for this one because it holds the full context of three or four competitor pages without collapsing the analysis.
You are a marketing strategist. I'm going to paste three competitor landing pages below. Compare them on:1. The single promise each page is making (in one sentence each)2. Which audience each is actually talking to (not who they claim to target, who they actually talk to)3. The most interesting word or phrase on each page — the thing that's specific to them4. Where each is weakest — a part of the page that could be attackedAt the end, give me one positioning gap I could claim that none of them are claiming yet. Make it specific, not a platitude.Competitor 1 ([NAME]): [PASTE or URL]Competitor 2 ([NAME]): [PASTE or URL]Competitor 3 ([NAME]): [PASTE or URL]
The gap-finding question is the payoff. Competitor analysis that doesn’t end in a gap is just a book report.
4. The 40-headline generator
For ad testing, email subject lines, hero copy. ChatGPT handles volume-plus-variety here better than Claude, which tends to converge on two or three “best” angles and repeat them.
Generate 40 headline variations for this product/offer. I want:- 10 in a direct-benefit frame (what you get)- 10 in a pain-avoidance frame (what you avoid)- 10 in a curiosity frame (open loop, no spoiler)- 10 in a status frame (what this says about the buyer)No more than 10 words each. No questions unless the question is genuinely surprising. Don't use "unlock," "transform," "revolutionize," "game-changer," "next-level," or any variant.Product/offer: [PASTE]Target audience: [PASTE]What they care about most: [PASTE]
Ban the slop words at the prompt level and you lose maybe 15% of the output to refusals, but the other 85% is usable. Without the ban list, you’re cleaning up “Unlock the transformation your business deserves” headlines for an hour.
5. The weekly email drafter
For the one email a week most operators keep meaning to send and don’t. This is ChatGPT territory — faster iteration, tighter voice match once you give it samples.
Draft a weekly email from me to my list. Constraints:- Under 250 words- First person, conversational, no "hope you're having a great week"- One single idea, not three- One specific story or example (I'll fill this in — leave a [BRACKET] if you need one)- Ends with one clear ask or one open question, not both- Subject line that would make sense out of context (no "weekly update" or "newsletter #12")Here are three of my past emails for voice reference: [PASTE]Here's the idea for this week: [PASTE]
The “no hope you’re having a great week” line is there because ChatGPT defaults to that opener roughly 70% of the time unless you block it. Same principle as the ad prompt — specify the failure mode, save yourself the cleanup.
6. The metrics read-out
For making sense of weekly numbers without hiring an analyst. Either tool works; I use whichever has my data already pasted into a longer thread.
I'm going to paste my weekly analytics pull below (GA4, newsletter, social, and pipeline). Tell me:1. What materially changed this week versus last week and the 4-week trailing average — and whether the change is noise or signal.2. Which single number I should pay the most attention to next week, and why.3. One thing that's weird or doesn't add up — something I should investigate before acting on the rest.Don't list every metric. Don't tell me things are "trending upward." Use plain numbers and plain verbs.Data: [PASTE CSV or table]
“Don’t tell me things are trending upward” kills a specific AI tic where every chart is described with the same three verbs. If the number went up by four percent, that’s usually noise. If it jumped thirty, that’s the headline.
7. The voice audit
For when drafts start sounding like AI and you can’t tell why. This is strictly Claude. ChatGPT struggles with self-audit because its default register is the thing we’re trying to audit out.
You are a brutal editor. I'm going to paste a draft below. Your job is to flag every sentence that sounds like AI wrote it. For each flag, give me:1. The exact sentence2. Why it reads as AI (specific pattern — be concrete, not "it's generic")3. A rewrite in the same voice as the rest of the pieceFocus on: pull-quote phrasing that's trying too hard, "not X — Y" parallelism, rule-of-three lists that don't need three items, vague significance claims ("this matters because..."), and throat-clearing openers.Draft: [PASTE]At the end, give the piece a 1-10 score on how human it reads, and name the single biggest fix.
I run every long piece through this at least once before publishing. It catches the drift you stop noticing after the tenth pass.
How to make these work harder
The prompts above are starting points. Three tweaks that turn a generic prompt into a brand-specific one:
Paste your style guide at the top of every prompt. Not a link — the actual text. Models can’t read URLs reliably. Pasting the guide into the system prompt (or the first message of every Claude Project / custom GPT) makes voice drift drop meaningfully.
Add a “don’t do” list. Not “avoid corporate jargon” — list the specific words, phrases, or opener patterns you see in the output that you want gone. The more specific, the more it sticks.
Give it three samples of your own past work. Not twenty — three. The model needs enough signal to pattern-match your voice without getting lost in variance.
Build a Claude Project or a Custom GPT for each of the 7 prompts above, loaded with your brand context, and you turn them from one-off utilities into a standing asset. By month three, you stop writing prompts from scratch and start iterating on the ones you’ve already trained.
A note on which model to use when
Claude is better for: longer context, judgment calls, editorial voice, anything that requires sitting with ambiguity. The 200k context window means a real discovery call plus a brand guide plus three draft versions all fit in one prompt.
ChatGPT is better for: volume, speed, image-adjacent work (headline + image together), quick variations, and anything where you’re going to run forty iterations and pick three.
You don’t have to pick one. The operators I know who ship the most marketing content have both open, and they’ve stopped thinking about which tool to use. They reach for whichever one is already in the right context.
If you try any of the above, swap the bracketed placeholders for your own and run them back-to-back on the same brief. The difference in output tells you more about each model’s personality than any benchmark will.
