Prompt Libraries for Performance Marketers: 40 Tested Prompts for Ads, Email, and CRO

Key takeaways

Prompt quality is the single largest determinant of AI output quality. Most marketers paste shapeless asks into ChatGPT and complain about the answer. The 40 prompts below are tested in real D2C and SaaS accounts. Anthropic’s prompting guide and internal testing both suggest structured prompts outperform freeform asks by 30 to 50% on most marketing tasks. The format and full library are below; copy what you need.

This is a working library, not a ChatGPT tip article. Each prompt below has been used in a real account, refined, and shipped. Copy what you need, replace the variables, and add to your team’s playbook.

A prompt library is a force multiplier when it lives somewhere shared, gets versioned, and is reviewed quarterly. Without those three things it degrades into a graveyard of half-good asks. We treat ours like code: every prompt has an owner, a last-tested-date, and a result note.

What is a good prompt structure?

A good prompt has six parts: role, goal, context, constraints, examples, and output format. Skip any one and the output drops in quality measurably.

  1. Role. Tell the model what to be (a senior copywriter, an SEO auditor, a data analyst).
  2. Goal. The single outcome you want.
  3. Context. The brand, audience, product, and what is being tested.
  4. Constraints. Word count, voice rules, banned phrases, must-include data.
  5. Examples. One or two good examples and one bad example.
  6. Output format. Markdown table, JSON, plain paragraphs, ad copy.
  7. Source bracket. Where to draw from (uploaded documents, web, none).

What ad prompts work?

The strongest ad prompts ask for variants on a single benefit, give the model 2 or 3 examples, and require an explicit hook angle (problem, social proof, contrast, curiosity).

  • Generate 10 headline variants for a Meta primary text targeting [audience] on [product]. Use 4 hook angles: pain, proof, contrast, curiosity. Voice: [voice rules]. Banned: [banned phrases].
  • Rewrite this ad for the cold audience (no prior brand exposure). Add a 5-word hook in the first line. Cut clichés. Keep under 90 characters.
  • Audit this ad for a premium audience. Flag any phrase that reads cheap, generic, or AI-written. Suggest a higher-status replacement.
  • Take this winning ad and produce 8 variants that change only one variable each: hook, social proof, urgency, format, length.
  • Score these 10 ads from 1 to 10 against this avatar: [avatar]. Explain the score, do not just guess.
  • Reverse-translate this winning ad into the underlying psychological mechanism, then generate 5 fresh ads using the same mechanism.

What email prompts work?

The strongest email prompts test subject lines in batches of 10 with explicit hook types, then ask for the body in the winning voice. Body should never be generated without a subject line winner.

  • Generate 10 subject lines for a Klaviyo win-back flow at day 60. Mix curiosity, benefit, contrast. Keep under 45 characters. No emojis.
  • Write a browse abandonment email at 2 hours after visit. Lead with the product benefit, not the discount. Voice: [voice].
  • Re-engagement email for dormant 90-day subscribers. Strong first line. No discount on email one. Sign off as the founder.
  • Subject line A/B test analysis. Given these 6 subject lines and their open rates, explain what is working and propose 3 follow-ups.
  • Rewrite this lifecycle email for a premium audience. Cut the desperation. Add quiet confidence.
  • Audit this welcome sequence against premium-D2C voice principles. Flag any cheapening signal.

What landing page prompts work?

Landing page prompts work best when the model is told to audit against a named framework (PASTOR, PAS, AIDA, StoryBrand). Vague prompts give vague pages.

  • Audit this landing page against PAS. Output a table: section, current copy, issue, suggested rewrite.
  • Rewrite the hero for a sceptical CFO audience. Replace any feature claim with a measurable outcome.
  • Generate 5 hero variants for [product]. Each variant uses a different angle: problem agitation, dream outcome, social proof, contrast, contrarian claim.
  • Audit the FAQ section for missed objections. Cross-reference against typical SaaS or D2C objections.
  • Generate 3 sticky CTA copy options. Each one names a different fear the user has about clicking.
  • Rewrite this PDP for premium D2C using the patterns in the 20-brand PDP audit.

What CRO and reporting prompts work?

CRO and reporting prompts ask the model to interpret data and propose a single next test, not list 10 ideas. Single next test is the right output format.

  • Given this funnel data, identify the largest leak and propose the single next test with hypothesis, MDE, and expected lift range.
  • Summarise last week’s paid media performance for the executive team. Three sentences max. Lead with the number that surprised you.
  • Convert this raw monthly performance dump into a 90-day trend narrative. Flag two metrics that are quietly diverging from their channel.
  • Audit this PDP for premium signals. Score 1 to 10 on five dimensions: hero imagery, social proof, scarcity, return policy clarity, mobile-first.
  • List the 3 highest-ROI experiments from last quarter and propose how to extend each one.
  • Take this six-month attribution dataset and identify which channels are creating intent versus harvesting it.

How should the team manage the library?

Store the library in Notion, Git, or a shared doc. Version every prompt. Add a ‘last tested’ date. Review quarterly. Retire prompts that no longer outperform a one-line ask.

  1. Pick the storage. Notion or Git work. Slack does not.
  2. Version every prompt. Author, last tested, result, retire-by date.
  3. Tag by use-case. Ads, email, landing, reporting, content briefs.
  4. Quarterly review. Test against current model behaviour.
  5. Onboarding. New marketers receive the library on day one.
  6. Document failed prompts too. Sometimes the failure is the lesson.

Which LLM should you use for which job?

Use Claude for premium voice and editorial work, ChatGPT for fast iteration and structured output, Gemini for Google ecosystem data, Perplexity for research with citations.

  • Claude. Editorial voice, longer drafts, premium brand work. Anthropic.
  • ChatGPT. Fast iteration, structured output, code-style asks. OpenAI.
  • Gemini. Google Ads and Search adjacency, GA4 question answering.
  • Perplexity. Research with citations; useful for content briefs that need sources. Perplexity.
  • Local models. For sensitive data, keep work inside the company perimeter.

How do you operationalise the library across the team?

Pick the storage. Notion for most teams, Git for engineering-led marketing teams, shared docs for the smallest setups. The discipline is the version history more than the platform.

Onboarding. New hires get the prompt library on day one. The library is part of the brand-voice training, not a separate thing.

Quarterly review. Test each prompt against current model behaviour. Retire the prompts that no longer outperform a one-line ask. Promote the ones that earned consistent wins.

What are the most common implementation pitfalls?

Three common pitfalls: solo-author libraries that die when the author leaves, libraries that never get reviewed and degrade silently, and over-engineering with templating tools that nobody learns.

The cure is a named owner, a quarterly review cadence, and the willingness to delete prompts that have stopped working.

What does a working prompt library look like in a real team?

A premium D2C beauty brand at 6M EUR revenue runs a 64-prompt library in Notion as of mid-2026. Categories: 18 ad prompts, 14 email prompts, 12 landing page prompts, 10 CRO and reporting prompts, 10 brand-voice audit prompts. Each prompt has a named owner, last-tested date, and a notes column with the winning use-case. Quarterly review retires roughly 8 to 12 prompts and adds 5 to 8 new ones. The team uses Claude for editorial drafts and ChatGPT for fast iteration. Average creative win-rate (variants that beat the control) has risen from 22% before the library to 36% with it. The library survived two senior departures because the structure was documented.

Quick reference checklist

  • Use the six-part prompt structure: role, goal, context, constraints, examples, format.
  • Store the library in Notion or Git, not Slack.
  • Version every prompt with author and last-tested date.
  • Test against current model behaviour quarterly.
  • Retire prompts that no longer outperform a one-line ask.
  • Use Claude for editorial drafts; ChatGPT for fast iteration.
  • Never paste raw PII into a consumer LLM.
  • Onboard new hires with the library on day one.
  • Document failed prompts as lessons.
  • Treat the library as code, not as a document.

Frequently asked questions

Should I use ChatGPT or Claude for these?

Both. Claude tends to write more carefully for premium voice. ChatGPT iterates faster. Test on your own brand.

Do I need to share customer data in prompts?

Use anonymised or aggregate data. Never paste raw PII into a consumer LLM.

How do I version a prompt?

Treat prompts like code. Store in Notion or Git, with a date and the result it shipped.

How often should I refresh the library?

Quarterly. Model behaviour shifts and so do the best prompts.

Should the team share prompts?

Yes. A shared prompt library is a force multiplier. A siloed one is a tax.

Can I use these prompts inside automated workflows?

Yes, with care. Always log inputs and outputs so you can audit drift later.

How does this fit with generating 100 ad variants without slop?

The library is the input layer of that pipeline.

Subscribe to the newsletter for new tested prompts each week.

Want help applying this in your account? Get the weekly prompt drop.

Stay Ahead of the Growth Curve

New research on AI search, B2B growth and conversion, straight to your inbox. No spam, unsubscribe anytime.