Skip Navigation
Writing AI Subject Lines Without Sounding Like a Generator
Email Copywriting

Writing AI Subject Lines Without Sounding Like a Generator

Models are good at volume and bad at your brand voice. Use them to draft options, then edit like a human who has to live with the unsubscribe rate.

by Mailcentr WizardAug 10, 20262 min read

AI can propose twenty subject lines in a second. Most of them share the same cadence, the same fake urgency, and the same empty adjectives. Readers have started to recognize that rhythm even when they cannot name it. Models are good at volume and bad at your brand voice. Use them to draft options, then edit like a human who has to live with the unsubscribe rate.

Give the model constraints

Paste three subject lines that already worked for you, the actual offer, the audience, and a ban list: unlock, game-changer, don’t miss, emoji spam, and clickbait that the body does not pay off. Ask for options that a specific person would send, not a brand voice document.

If you only paste “write subject lines for our newsletter,” you will get the same newsletter the rest of the internet is generating. Garbage in is still the rule.

Edit for specificity

Replace vague benefits with a number, a date, or a concrete object. “Your March invoice is ready” beats “Important account update”. If the model cannot cite a detail from your brief, the line is filler. Specificity is also a deliverability feature: mail that matches what the person asked for gets fewer complaints.

Never skip the human pass

Read the line next to the preview text and the first sentence of the body. If all three repeat the same phrase, cut two of them. Run a small A/B test against a human-written control. Keep a swipe file of winners so the model is not your only memory.

Watch for invented facts. Models will put a discount or a date in the subject that the body does not contain. That is not clever. That is a complaint.

Process that survives a bad week

Generate ten, kill seven in sixty seconds, edit three, ship one against a control. Log the hypothesis. If the generator wins, it still went through a person. If it loses, you learned something about your list that a prompt library will not tell you.

Brand voice is a pile of winners

Do not maintain a 40-page voice PDF the model will ignore. Maintain twenty subject lines that shipped and did not get you hated. That file is the prompt. Update it when a line wins a real test.