AI 電子郵件自動化：如何自動撰寫、寄送與最佳化電子郵件

The four layers are: writing (using Claude with a structured prompt and email marketing skill file to produce subject lines, preview text, and body copy ready for light editing on the first run); sending intelligently (AI send time optimisation analysing each subscriber's historical open behaviour to send at their individual predicted optimal moment, consistently improving open rates 10–20% versus fixed send times); personalisation (ranging from merge tags through segment-based variants to dynamic content blocks and AI product recommendations, each layer adding implementation complexity and revenue impact); and optimisation (a monthly audit cycle identifying underperforming emails, diagnosing them with Claude, A/B testing the improved variant, and documenting what changed — compounding improvement over time).

The structure that consistently produces ready-to-edit output: specify the email type and audience segment; define the conversion goal precisely; request five subject line options each using a different psychological mechanism (curiosity, urgency, benefit, social proof, direct); request three preview text options under 90 characters; specify word count with instructions that the opening line must create curiosity or confirm the reader is in the right place and that there should be one CTA only; state the brand tone; and explicitly list what the email must not do — starting with pleasantries, using exclamation marks, or leading with the company name. Loading a brand voice skill file before running this prompt improves output quality further.

Use AI for the high-volume, repeatable formats: nurture sequences, transactional templates, re-engagement campaigns, newsletter drafts, A/B test variants, and seasonal campaigns. Write manually with AI polish for situations where the personal relationship is the point: critical launch emails, personal messages from a named sender, crisis communications, and highly relationship-dependent outreach where the authenticity of a human voice matters to the recipient. The distinction is between emails where quality and volume are the goal versus emails where a specific human relationship is on the line.

Three platform implementations deliver consistent results: Klaviyo Smart Send Time analyses each contact's historical open behaviour and sends within a defined window when that individual is most likely to open, producing 10–20% open rate improvements in tests versus fixed send times. HubSpot Send Time Optimisation works on the same principle but requires 90 or more days of contact send history before becoming fully predictive. ActiveCampaign Predictive Sending functions similarly and works best in nurture sequences where timing momentum across the sequence matters for engagement. All three require sufficient historical send data per contact — below that threshold they default to statistical averages rather than individual prediction.

Four steps taking under an hour total: a 30-minute performance audit exporting open rate, click rate, and conversion rate for every active automation email and flagging anything performing more than 20% below benchmark. A 15-minute Claude diagnosis session pasting the underperforming email plus its performance data and asking specifically what to change in the subject line, opening line, and CTA and why. Implementing the suggested changes as a B variant and running an A/B test for two weeks before applying the winner. Documenting what changed and how performance shifted — over six months this builds a pattern library showing which email structures work for your specific audience, compounding improvement with every cycle.