AI Prompts for Marketers: 10 Templates That Actually Convert (2026)

Marketing copy fails when the model writes for 'everyone' — which means no one. The fix is to load the prompt with specifics: who the reader is, what they already believe, what they're afraid of, and the single action you want. Three inputs do most of the work: a concrete audience (not 'small businesses' but 'solo bookkeepers who hate QuickBooks'), the offer and its one differentiator, and the channel's constraints (character limits, tone, platform norms).

Give the model one or two examples of copy you like and it will match the register far more reliably — few-shot prompting, per Brown et al., 2020. And always ask for variants: the first option is rarely the best, and seeing three lets you pick the angle that fits.

These prompts run well on any current mid-tier or frontier model. For high-volume copy at low cost, Gemini 3.5 Flash (~$1.50 in / $9.00 out per 1M) or gpt-5.4-mini ($0.75 / $4.50) are economical; for flagship brand copy where craft matters, Claude Opus 4.8 or gpt-5.5 produce noticeably better long-form. Prices as of June 2026 (OpenAI, Gemini).

When to use: paid social or search ads where you need angles to test, not one safe option.

``` Write 5 ad variations for [PLATFORM, e.g. Meta / Google Search]. Product: [WHAT IT IS] Audience: [SPECIFIC reader — role, context, what they currently use] One differentiator: [the single thing only we do well] Desired action: [click / sign up / buy] Each variation must use a different angle: pain-point, outcome, social-proof, curiosity, objection-handling. For each, give a headline (max [N] chars), primary text, and the angle name. No hype words ("revolutionary", "game-changer"), no exclamation points. Lead with the reader's problem, not our product name. ```

Why it works: forcing five distinct angles gives you a real test matrix instead of five rewordings of the same idea. Banning hype words and product-first openers is what makes the copy sound like a person rather than a press release. Build the structured version with the Ad Copy Generator.

When to use: you have a post or asset and need captions tuned to each channel's norms.

``` Write a caption for [PLATFORM]. Match the platform's norms: [Instagram = visual hook + line breaks + 3-5 niche hashtags; LinkedIn = a hook line, a short story, a takeaway, no hashtag spam; X = under 280 chars, one idea]. Post is about: [TOPIC] Goal: [save / share / click / comment] Voice: [describe brand voice in 3 words] Give me 3 options with different opening hooks. The first line must earn the second line. No "In today's world" openers. ```

Why it works: each platform rewards a different shape, so naming the norms per channel stops the model from posting a LinkedIn essay to X. "The first line must earn the second line" is the single most useful caption constraint — it forces a real hook. The Social Media Caption Generator does this with channel presets.

When to use: you're publishing a page and want titles and meta descriptions that earn the click without keyword-stuffing.

``` Write 5 title tags and 5 meta descriptions for a page about [TOPIC] targeting the search intent: [informational / commercial / transactional]. Primary keyword: [KEYWORD] Rules: - Title: under 60 characters, keyword near the front, one clear benefit. - Meta description: 150-160 characters, includes the keyword naturally, ends with a reason to click. No clickbait the page can't deliver. - Each option should target a slightly different angle or modifier. Return as a table: Title | char count | Meta | char count | angle. ```

Why it works: the character-count column keeps you inside what Google actually renders, and tying each option to an angle gives you variety to test in Search Console. "No clickbait the page can't deliver" protects your bounce rate. Generate these directly in the SEO Meta Generator.

When to use: any campaign or newsletter send where open rate is the bottleneck.

``` Write 8 subject lines for an email to [AUDIENCE] about [TOPIC/OFFER]. Mix the approaches: curiosity, benefit, urgency (only if real), question, and plain/specific. Keep each under [40-50] characters so it doesn't truncate on mobile. For each subject line, write matching preview text (35-90 chars) that extends it rather than repeating it. Flag any that might trip spam filters (ALL CAPS, "free", excessive punctuation). ```

Why it works: pairing each subject with preview text that extends rather than repeats it is what most people miss — the two lines are one unit in the inbox. The length cap prevents mobile truncation, and the spam-flag step saves a deliverability headache.

When to use: you need hero, benefits, and CTA copy that all stay consistent with one positioning.

``` Write landing page copy. First, restate the positioning back to me in one sentence so I can confirm it before you continue. Product: [WHAT IT IS] For: [AUDIENCE] who struggle with [PROBLEM] Unlike: [ALTERNATIVE], we [DIFFERENTIATOR] Proof: [any real proof points — numbers, integrations, guarantees] Then produce: hero headline + subhead, 3 benefit blocks (benefit-first, feature second), one objection-handling section, and a CTA. Benefits must be outcomes for the reader, not features of the product. ```

Why it works: asking the model to restate the positioning first catches a wrong premise before it writes 400 words on it. The 'benefit, then feature' rule is the difference between copy that sells and a spec sheet.

When to use: before writing anything, to pressure-test who you're actually talking to.

``` Act as a skeptical strategist. For this audience: [DESCRIBE], help me map: 1. The job they're trying to get done (in their words, not marketing). 2. Their top 3 pains and the workaround they use today. 3. The objection most likely to stop them buying [OFFER]. 4. The proof that would overcome that objection. 5. Three message angles, ranked by likely resonance, with a one-line rationale each. Where you're inferring rather than certain, say so. Do not invent statistics or survey data. ```

Why it works: the 'in their words, not marketing' instruction surfaces real language you can lift into copy. The 'do not invent statistics' guard matters — models will happily fabricate a market-size number if you let them. Use this output to seed a customer persona or brand voice.

When to use: you wrote one good thing (a post, talk, or article) and want a week of content from it.

``` Turn the source below into a content set: - 1 LinkedIn post (hook + insight + takeaway) - 3 X posts (one idea each, standalone) - 1 short email teaser linking back - 5 quote/stat cards (pull only claims actually in the source) Keep the core argument intact. Do not add claims, numbers, or examples that aren't in the source. Match the voice: [3 words]. Source: [PASTE] ```

Why it works: 'pull only claims actually in the source' is what keeps repurposing honest — otherwise the model embellishes your stat cards with numbers you never said. One source, many shapes, same argument. Plan the calendar with the Content Calendar Generator.

When to use: ecommerce or catalog copy where every competitor's description reads the same.

``` Write a product description for [PRODUCT]. Structure: one-line hook, a short paragraph on the experience of using it (not a feature dump), then 3-4 bullet specs. Write for [AUDIENCE]. Lead with the outcome they want. Avoid the clichés every listing uses: "premium quality", "perfect for", "elevate your", "must-have". If a claim needs proof, mark it [VERIFY]. ```

Why it works: banning the four clichés that saturate ecommerce copy immediately differentiates the listing. The [VERIFY] tag flags claims you need to substantiate before publishing. See the Product Description Generator for the templated version.

When to use: you want to run a real test, not just 'try a different button color.'

``` Design an A/B test for [ELEMENT — headline / CTA / email subject]. 1. State a clear hypothesis: "If we change X to Y, [metric] will improve because [reason about the reader]." 2. Write the control (current) and the variant copy. 3. Name the single metric that decides the winner. 4. Note the smallest meaningful difference worth shipping. Current copy: [PASTE] What we believe about the audience: [NOTE] ```

Why it works: forcing a falsifiable hypothesis before the copy stops you from testing random changes you can't learn from. Naming one deciding metric prevents the 'it improved clicks but hurt conversions' ambiguity.

When to use: you have copy you like and want a reusable voice spec the whole team (and the model) can follow.

``` From the samples below, extract a brand voice guide: - 3-5 voice traits, each with a one-line definition - "We sound like / we never sound like" pairs - 5 words we use and 5 we avoid - A before/after rewrite showing a generic sentence in our voice Base it only on the samples. Do not invent traits the copy doesn't show. Samples: [PASTE 3-5 examples] ```

Why it works: a voice guide grounded in real samples gives you a spec you can paste into every future prompt for instant consistency. 'Base it only on the samples' keeps it from inventing an aspirational voice you don't actually use. The Brand Voice Generator turns this into a reusable profile.

For high-volume, lower-stakes copy (captions, product descriptions, subject-line batches), an efficiency-tier model is plenty: gpt-5.4-mini ($0.75 in / $4.50 out per 1M) or Gemini 3.1 Flash-Lite ($0.25 / $1.50). For flagship landing pages, positioning, and long-form where craft is judged by humans, a frontier model — Claude Opus 4.8 ($5 / $25) or gpt-5.5 ($5 / $30) — produces noticeably better cohesion and specificity. Whatever you use, the input quality (real audience, real offer, real proof) matters more than the model. Prices as of June 2026; verify current rate cards (OpenAI, Anthropic, Gemini).

Sources and further reading: DAIR.ai Prompt Engineering Guide, Learn Prompting, Few-shot prompting (Brown et al., 2020), Gemini prompting strategies. Pricing current as of June 2026.