Case Study: Alex's Campaign Copy Crisis — A Three-Round Repair
The Deadline and the Disaster
It was Thursday morning, two days before Brightleaf's Black Friday promotional email was scheduled to send. The brief was simple: a promotional email announcing 30% off all products, targeting existing customers, with the goal of driving purchase before the holiday weekend. Target list: 45,000 subscribers. Subject line A/B test already set up.
Alex needed the email body copy. She had used AI tools for copy for months at this point and had a well-developed few-shot pattern library. But Thursday was slammed — back-to-back calls until noon, a brand review at 1 PM. She was working fast, and she made three mistakes.
She got three versions of the same fundamental problem. Here's the diagnostic transcript of what went wrong and how she fixed it.
First Attempt: The Corporate Email
Alex's first prompt (rushed, no pattern used):
"Write a Black Friday promotional email for Brightleaf. 30% off all products. Existing customers. We want them to buy before the weekend."
AI Output (Version 1):
Subject: Black Friday Sale — 30% Off Everything at Brightleaf
Dear Valued Customer,
We are thrilled to announce our Black Friday sale event, offering an exceptional 30% discount on our entire product range. At Brightleaf, we remain committed to providing you with sustainable, high-quality products that support your eco-conscious lifestyle.
This limited-time offer represents a significant opportunity to stock up on your favorite Brightleaf items or discover new products that align with your values. Whether you're looking for kitchen essentials, personal care products, or household solutions, our comprehensive range has something for everyone.
To take advantage of this offer, simply use code BLACKFRIDAY30 at checkout.
Best regards, The Brightleaf Team
Offer valid November 29–30. Terms and conditions apply.
Alex's immediate reaction: "This sounds like a press release from 1998. 'We are thrilled to announce.' 'Valued Customer.' 'This limited-time offer represents a significant opportunity.' None of this sounds like Brightleaf."
Diagnosis: Round 1
Alex took 2 minutes to diagnose before writing the repair prompt.
Specific failure: Wrong brand voice — corporate, stiff, generic. This email could have come from any retail brand. It has none of Brightleaf's warmth, informality, or light humor.
Root causes identified: 1. Insufficient context (Primary): She didn't include her brand voice reference examples. The AI had no idea how Brightleaf actually talks. 2. Format mismatch (Secondary): "Dear Valued Customer" isn't how Brightleaf addresses its audience. The structure is also wrong — a list-like body rather than the hook → benefit → CTA structure she uses.
Triage Matrix: Low distance from goal on structure (the basic elements are there — sale announcement, offer, CTA), but the brand voice failure is a complete miss. Distance: 3. Repair effort: Medium (reloading the brand voice context is the main work). Decision: Repair.
Repair Prompt 1: The Context Reload
This email doesn't sound like Brightleaf at all. I need to reload the brand context I forgot to include.
Brand voice reference — please match this style exactly:
---
[Example 1: Bamboo cutting board product description]
"Your kitchen's hardest worker, quietly doing its thing since day one..."
---
[Example 2: Email campaign header for reusable bags]
"We have a theory: the best swaps are the ones you barely notice..."
---
[Example 3: Instagram caption for dish soap]
"One bottle. Six refills. Math that makes sense..."
---
Email structure for Brightleaf campaigns:
- No "Dear Valued Customer" — address as "you" informally
- Hook (first 1-2 sentences that make someone want to keep reading)
- The offer stated simply and directly
- One specific benefit statement (not a list)
- Clear CTA with the code
- Short — 3 brief paragraphs max
Now please rewrite the Black Friday email with this voice and structure.
AI Output (Version 2):
Subject: The best kind of Friday
It's Black Friday. You know what that means — but we're guessing you weren't expecting to save 30% on your dish soap, your tote bag, and the bamboo cutting board you've been eying since summer.
30% off everything. No exclusions. No fine print except the usual: offer ends Sunday, code BLACKFRIDAY30 at checkout.
Stock up on the things you use every day, or finally grab that thing you kept almost buying. Either way, your cart's going to look good.
Alex's reaction: "Better. Much better. This sounds like Brightleaf. But there are two problems."
Diagnosis: Round 2
Specific failure 1: "The bamboo cutting board you've been eying since summer" — this is oddly specific. It assumes all customers have been eyeing a specific product, which is awkward and could feel presumptuous to customers who have never seen a bamboo cutting board.
Specific failure 2: The offer mechanics are incomplete. "Offer ends Sunday" — but Sunday is not the end of the Black Friday window. Brightleaf's sale was running Thursday through Monday. This is factually wrong.
Root causes: 1. Hallucination (light form): "The bamboo cutting board you've been eying since summer" is the model assuming/inventing a specific customer behavior that Alex never specified. 2. Insufficient context: She didn't specify the exact offer dates.
Triage Matrix: The voice and structure are now right (distance from goal: 2). Two targeted fixes needed. Repair.
Repair Prompt 2: Two Targeted Corrections
Almost right. Two specific corrections needed:
1. "The bamboo cutting board you've been eying since summer" is too product-specific
and presumptuous. Replace this with something that speaks to the range broadly —
something they use every day or have been meaning to add to the cart.
Keep the same casual, warm tone.
2. The offer dates are wrong. The sale runs Thursday November 28 through
Monday December 2 (5 days, not just the weekend). Please update the
"offer ends Sunday" line to reflect this accurately.
Keep everything else exactly as is — the subject line, the voice, the structure.
AI Output (Version 3):
Subject: The best kind of Friday
It's Black Friday. You know what that means — but we're guessing you weren't expecting to save 30% on your dish soap, your tote bag, and the kitchen staples you actually use every week.
30% off everything. No exclusions. No fine print except the usual: offer ends Monday, December 2, code BLACKFRIDAY30 at checkout.
Stock up on the things you use every day, or finally grab that thing you kept almost buying. Either way, your cart's going to look good.
Alex's final review:
She read it three times. It sounded right. The subject line was good. The offer was accurately stated. The voice matched Brightleaf. The CTA was clear.
She forwarded it to the brand team for a final read. One minor suggestion came back: change "your cart's going to look good" to "your cart will thank you" — a Brightleaf-ism the brand team preferred. Two-minute edit.
The email went out Friday morning. Open rate: 34.2% (vs. the campaign average of 28%). Click-through rate: 11.8% (vs. average 8.4%). Black Friday revenue exceeded target by 17%.
The Repair Transcript Analysis
Looking back at the three rounds:
What Alex got wrong in Round 1: She skipped her pattern — she was working fast and didn't include the brand voice reference examples. This is the most common way her patterns fail: she forgets to use them. The diagnostic framework made this visible immediately.
What she diagnosed correctly: The primary root cause was insufficient context, not the model's failure. This is the key diagnostic insight: when an output is generic and off-brand, the fault is almost always missing context, not model incapability.
What the repair prompts did right: - Repair 1 was a context reload with the full few-shot reference. It didn't try to describe the voice — it demonstrated it. - Repair 2 was surgical: "Keep everything else exactly as is" prevented the model from changing the elements that were now correct. - The specificity of "The bamboo cutting board you've been eying since summer" is too presumptuous and product-specific — rather than vague feedback ("it's too specific"), the feedback quoted the exact problem text.
Total time for all three rounds: 19 minutes, including reading outputs and writing repair prompts. A full manual rewrite from scratch would have taken 45-60 minutes.
What Alex Learned
Lesson 1: Using a pattern is only valuable if you actually use it. The brand voice reference library existed in her pattern library. She just didn't include it. The most sophisticated pattern library in the world doesn't help if you're working fast and skip the template. She added a personal rule: when generating any customer-facing copy, pattern loading is non-negotiable, even under time pressure.
Lesson 2: Specific failure identification produces specific repairs. When she wrote "this sounds corporate," that was too vague to repair from. When she said "the bamboo cutting board you've been eying since summer is too product-specific and presumptuous," that produced an exact fix. The more specifically you name the failure, the more targeted and effective the repair.
Lesson 3: "Keep everything else as is" is the most important instruction in a targeted repair prompt. Without it, repair prompts can fix the stated problem while introducing new problems. "Keep everything else exactly as is" limits the repair scope.
Lesson 4: Verify factual claims even in marketing copy. The dates were wrong because she didn't specify them. In a promotional email, wrong dates are a customer experience problem that generates complaints and erodes trust. Any time-sensitive factual element in a prompt output — dates, codes, prices, product specifications — should be explicitly verified.
Prevention list for future email prompts: 1. Always load the brand voice reference from the pattern library (non-negotiable) 2. Include exact dates, codes, and prices in the brief — don't let the model invent them 3. Specify the email structure explicitly, not just "an email" 4. After any AI draft: read specifically for assumed product details and off-brand phrases
These four prevention items are now in her email campaign pattern documentation under "Notes — Common Failure Modes."