Generate A/B test variants of your best cold email
Turns your control email into a disciplined test plan: a diagnosis of its current mechanics, four variants that each isolate one variable — angle, proof, CTA, or opener — and a ranked prediction of which to test first. It also checks whether your send volume can even power a meaningful result.
You are a cold email testing strategist. You've run enough experiments in Smartlead and Instantly to know why most A/B tests teach nothing: people test trivia (comma placement, greeting choice) or change five things at once and can't attribute the result. Real tests isolate one high-leverage variable — the angle, the proof type, the CTA friction, or the opener style — and hold everything else still. I'm going to paste my control email. Do this: 1. DIAGNOSE: Identify the email's core angle, proof type, CTA style, and opener mechanism in one line each. This is what we're testing against. 2. GENERATE 4 VARIANTS, each changing exactly ONE variable and labeled with its hypothesis: - Variant A: different ANGLE (e.g., cost framing swapped for risk framing) — same structure, proof, and CTA. - Variant B: different PROOF TYPE (number swapped for named customer, or case detail for social proof) — everything else held. - Variant C: different CTA FRICTION (interest question swapped for resource offer, or vice versa) — everything else held. - Variant D: different OPENER MECHANISM (observation swapped for question, or problem statement for data point) — everything else held. 3. PREDICT: Rank the variants by expected lift for my audience, with one sentence of reasoning each, and tell me which ONE to test first — testing all four at once splits samples too thin at normal volumes. Keep every variant within 10 words of the control's length so length never confounds the test. Before you write anything, interview me. Ask me these questions ONE AT A TIME, waiting for my answer each time: 1. Paste the control email, subject line included. 2. What's it performing at, over how many sends? 3. Who's the audience, and what's your monthly send volume to them? 4. What have you already tested, and what happened? Once you have my answers, produce the diagnosis, variants, and test plan. If my volume can't power a meaningful test, say so and recommend the minimum sample before I bother.
How to use it
- 1
Copy the prompt into Claude, ChatGPT, or any LLM.
- 2
Paste your genuine control with its real performance numbers — the test plan calibrates to both.
- 3
Run the recommended variant against the control at 50/50 in your sequencer, minimum a few hundred sends per arm.
- 4
Return with results; the losing hypothesis is information, and the next variant in the ranking is your next test.
Best practices
One variable per test, always — the labeled hypothesis exists so you know what you learned when the numbers come in.
Judge by reply rate or positive-reply rate, not opens; open tracking is unreliable and opens don't book meetings.
Keep a test log of hypothesis, dates, and result per variant — three months of logged tests beats any guru's template advice.
When a variant wins decisively, it becomes the new control and the cycle restarts from the diagnosis.
Example: what this looks like in practice
A RevOps manager at a contract management platform has a control pulling 2.9% replies over 1,400 sends to legal ops leaders. The diagnosis labels it: cost-saving angle, ROI-number proof, demo-ask CTA, observation opener. The model predicts the CTA variant first — the demo ask is high friction for a first touch — and swaps it for 'worth seeing how the redline comparison works?'. At 400 sends per arm, the variant hits 4.4% against the control's 3.0%. Variant C becomes the new control, and the angle test is queued next, per the ranking.
Best fit
This prompt is one gear in a bigger machine. We orchestrate 20+ tools into outbound systems our clients own — and guarantee the results.
Apply for a Pilot Spot → →Frequently asked questions
Change exactly one variable per test — angle, proof type, CTA friction, or opener style — hold everything else constant including length, split your list randomly, and judge by reply rate over at least a few hundred sends per variant. Most 'tests' fail by changing several things at once, which produces a winner but no lesson.
More cold email writing prompts
Get your cold email critiqued and scored like a skeptical prospect
Runs your cold email through a three-pass review: a brutal line-by-line read from a busy prospect's perspective, a five-dimension strategic score, and a rewrite that fixes the weakest dimensions. You learn why the email underperforms, not just that it does — the pattern transfers to every email you write after.
Write a complete cold email from scratch
Produces a complete, send-ready cold email of 60-100 words with the four-part spine that books meetings: reason, problem, proven outcome, soft ask. The five-question interview forces you to articulate the 'why now' most campaigns skip, and the prompt refuses to write until that angle is real.
Audit and rank your subject lines before you hit send
Puts every subject line candidate through a brutal pre-send audit: a 1-10 plausibility score, its biggest weakness named, spam-pattern flags, and a fixed rewrite. You get a ranked list with a clear verdict — two to test, the rest to kill — instead of sending on gut feel.