The Exact Prompts We Use to Generate Winning Campaigns Weekly
The Exact Prompts We Use to Generate Winning Campaigns Weekly
Most companies talk about using AI for marketing. Very few show you the actual prompt. That gap between "we leverage AI" and "here is the 400-word instruction block that produced our 340% ROAS email sequence" is where most of the value lives.
This article walks through five prompt frameworks we run every single week to generate campaigns across email, paid social, and landing pages. These are not toy examples. They are production-grade structures, adapted from weeks of A/B testing, that consistently outperform our previous human-only workflow.
The Meta-Prompt: Why Structure Beats Creativity Here
The single biggest mistake teams make when prompting for marketing copy is writing a single generic prompt like "Write an email about our summer sale." The output is competent and useless. It reads like a template because it is one.
The prompts below all share a common architecture:
Role assignment — narrow the AI's behavioral surface
Context injection — feed it real data, not vibes
Constraint stacking — force specificity through rules
Output scaffolding — define the exact shape of the response
Iteration hooks — build in a mechanism for the AI to critique and refine its own output before returning it
You do not need a new model. You need a new prompt shape.
Prompt 1: The Audience-Split Email Sequence
Use case: Weekly drip email (3-part nurture)
Prompt:
ROLE: You are a senior lifecycle email strategist at a DTC brand
selling [PRODUCT CATEGORY] to [DEMOGRAPHIC]. You have written
40+ sequences that exceeded 52% open rates and 11% click-through.
CONTEXT:
- Product: [PRODUCT NAME], [PRICE POINT], one-line value prop: [X]
- Audience segment: [e.g., "abandoned cart, viewed 2x, no purchase
in 45 days, median LTV of $87"]
- Last email they received: [SUBJECT LINE + 1-SENTENCE SUMMARY]
- Brand voice tokens: [e.g., "confident but not salesy, short
paragraphs, zero exclamation marks, conversational contractions"]
CONSTRAINTS:
- Three emails, spaced Day 1 / Day 3 / Day 6
- Each email: subject (≤ 45 chars), body (120–180 words), CTA button text
- Email 1: reframe the problem, no product mention in first 80 words
- Email 2: social proof + objection handling, reference a specific
use case
- Email 3: urgency with a concrete mechanism (not "hurry"), state
what happens if they do nothing
- No clichés: "in today's world," "game-changer," "unlock"
OUTPUT FORMAT:
Return a JSON object:
{
"email_1": {"subject": "...", "body": "...", "cta": "..."},
"email_2": {"subject": "...", "body": "...", "cta": "..."},
"email_3": {"subject": "...", "body": "...", "cta": "..."}
}
SELF-CRITIQUE (append after JSON):
List the 2 weakest lines across all three emails and rewrite them
in place.Why this works weekly: You swap the [PRODUCT CATEGORY], [DEMOGRAPHIC], and [AUDIENCE SEGMENT] fields every cycle. The structure stays identical, so the output is consistent enough to A/B test against last week's version. We typically run two prompt variants (same structure, different constraint on tone) and let the data pick a winner.
Prompt 2: The Ad Creative Matrix
Use case: Paid social (Meta, TikTok, LinkedIn) ad copy — 12 variants per week
Prompt:
ROLE: You are a paid social copywriter who has spent $2M+ in
creative testing. You think in hooks, not paragraphs.
CONTEXT:
- Ad platform: [PLATFORM]
- Format: [e.g., "15-second UGC-style video script" or "static
image + 2-line primary text"]
- Product benefit (one sentence, no jargon): [X]
- Target persona: [FIRST NAME], [AGE RANGE], [JOB TITLE OR ROLE],
their #1 frustration with current solutions: [X]
- Offer: [e.g., "free 14-day trial, no card required"]
- Competitor angles to AVOID (we are not them): [LIST 2-3]
CONSTRAINTS:
- Generate 12 distinct ad concepts
- Each must open with a pattern interrupt in the first 3 words
(video) or first line (text)
- Hook types to distribute: contrarian, question, specific number,
story fragment, mistake callout, "you're wrong" reframe
- Body: max 2 sentences before the CTA
- CTA: verb-first, no "learn more" or "discover"
- Banned words: leverage, synergy, solution, journey, unlock,
elevate, seamless
OUTPUT FORMAT:
| # | Hook Type | Hook (first 3 words / line) | Body (≤2 sentences) | CTA |
|---|-----------|------------------------------|----------------------|-----|
SELF-CRITIQUE:
Score each concept 1–10 on: (a) scroll-stop probability, (b)
brand safety, (c) CTA friction. Flag any scoring below 7 and
rewrite.Why this works weekly: The 12-variant matrix gives the media buyer enough volume to find a winner without waiting for next week. The hook-type distribution constraint prevents the AI from clustering five ads around the same angle, which is what happens without it.
Prompt 3: The Landing Page Section Builder
Use case: Weekly A/B test on a hero section or offer module
Prompt:
ROLE: You are a conversion copywriter specialized in [INDUSTRY]
landing pages. Your work has been tested on pages with > 50K
monthly sessions.
CONTEXT:
- Current hero headline: [X]
- Current subhead: [X]
- Current CTA: [X]
- Current conversion rate: [X]%
- Target customer's #1 hesitation before clicking CTA: [X]
- Three proof points available: [1, 2, 3]
- Brand constraints: [e.g., "no more than 30 words in headline,
benefit-first not feature-first"]
TASK:
Generate 5 hero section variants. For each:
- Headline (≤ 30 words, benefit-led)
- Subhead (1 sentence, addresses the hesitation directly)
- CTA button text (≤ 5 words)
- One-line "proof strip" beneath CTA (e.g., "Trusted by 4,200+
teams" or "No. 1 rated on [PLATFORM]")
CONSTRAINTS:
- No two variants may use the same psychological trigger.
Triggers available: loss aversion, social proof, specificity,
identity, curiosity gap, effort contrast.
- At least one variant must invert the expected structure
(proof before headline, or CTA before subhead)
OUTPUT FORMAT:
For each variant, a block:
VARIANT [N]
Headline: ...
Subhead: ...
CTA: ...
Proof strip: ...
Trigger used: ...
SELF-CRITIQUE:
Rank all 5 by predicted lift vs. current version (express as a
range, e.g., "+8–12%"). Justify the top pick in 2 sentences.Prompt 4: The Objection-Killing FAQ Generator
Use case: Weekly refresh of sales-page FAQ or support-deflection content
Prompt:
ROLE: You are a customer success lead who has read 2,000 support
tickets for [PRODUCT]. You know every objection before it is
spoken.
CONTEXT:
- Product: [X]
- Top 5 reasons prospects do NOT buy (from call recordings /
sales CRM notes): [LIST]
- Current FAQ has [N] items. Three weakest (lowest helpfulness
rating from users): [LIST WITH REASONS]
TASK:
Rewrite or generate FAQ entries that directly neutralize the top 5
non-purchase reasons. For each:
- Question (phrased exactly as the customer would say it, informal)
- Answer (3–5 sentences, acknowledge the concern in sentence 1,
then resolve it with a specific fact or example)
- "But what if..." follow-up (address the second-order objection
that appears after the first answer)
CONSTRAINTS:
- No hedging language: "generally," "in most cases," "it depends"
- One specific number, statistic, or example per answer
- Tone: helpful colleague, not corporate bot
OUTPUT FORMAT:
Q: ...
A: ...
But what if...: ...
SELF-CRITIQUE:
For each answer, identify if it would survive a skeptical customer
asking "say that differently, I don't believe you." If not,
rewrite.Prompt 5: The Weekly Performance Retro-Feed
Use case: Close the loop — feed last week's data back into next week's prompt
Prompt (run Friday, output feeds Monday's campaigns):
ROLE: You are a growth marketing analyst. Your job is not to
describe what happened. Your job is to extract the ONE variable
that changed outcomes so we can lock it in.
CONTEXT:
- Last week's campaigns: [PASTE SUMMARY — channel, spend, CTR,
CVR, ROAS per variant]
- Hypothesis we were testing: [X]
- Segment that overperformed vs. control: [X]
- Segment that underperformed: [X]
- Creative fatigue indicator: [e.g., "Ad B frequency hit 3.2 on
Day 5"]
TASK:
1. Winner: Which single variable most likely drove the outperformance?
Name it in one phrase.
2. Loser: Which variable most likely dragged performance? Name it.
3. Constraint for next week's prompt: Write a 1–2 sentence
instruction I can paste into the context block of next week's
campaign prompt that bakes in this learning.
4. Kill or keep: For each underperforming variant, state in 5 words
or fewer why it died.
OUTPUT FORMAT:
Winner: [phrase]
Loser: [phrase]
Next week's constraint: "[pasteable instruction]"
Kill/Keep:
- Variant A: ...
- Variant B: ...Operationalizing the System
None of this works as a one-shot. The weekly cadence looks like this:
Day | Action | Output |
|---|---|---|
Mon AM | Run Prompt 1 + 2 with fresh segment data | Email sequence + 12 ad concepts |
Mon PM | Run Prompt 3 on top-performing landing page variant | 5 hero variants queued for A/B |
Tue | Ship campaigns. Assign variant IDs. | Live experiments |
Thu | Run Prompt 4 on any product/offer change | Updated FAQ deployed |
Fri | Run Prompt 5. Paste output into Monday's prompt context. | Refined constraints |
The prompts compound. Week 4 is not the same as Week 1, because the constraint blocks have been tightened by real performance data. The AI is not getting smarter. Your prompt is. That is the actual leverage.
Start with one prompt. Run it for three weeks before adding another. The system beats the single heroic prompt every time.
