Best AI Prompts for Email Writing

Why do AI-written emails get ignored?

The 1.7% reply rate is not a hard ceiling. It is the average of a population that includes a lot of AI-pattern-matching emails. The senders who land 6-10% reply rates are the ones writing emails that read as personal, specific, and short. Gmail’s spam filtering in 2026 also factors AI-pattern signals: messages with default ChatGPT openings, em-dashes in every sentence, and rhetorical-question closes get a quiet deliverability penalty that does not show up as a bounce but shows up as silence.

The biggest predictor of reply rate in cold outreach data is not subject-line cleverness or send time. It is sentence count. Under 5 sentences beats over 10 by a factor of 3 to 4.

Apollo and Hubspot, 2025 outbound benchmarks

The eight prompts below cover the eight most common email situations. Each one is short. Each one names what to keep specific and what to cut. Each one assumes you will edit the output before you send. Sending the first draft of any of these is the failure mode that produces the 1.7% baseline.

What are the eight email prompts?

1. The cold outreach email

Why this works: The 5-sentence and 6-word ceilings force concision. Three versions force comparison. The “one specific thing not from their company website” is what separates a researched email from spray-and-pray. The yes/no question close is the format with the highest measured reply rate in 2025 outreach data.

2. The follow-up to a non-reply

Why this works: The forbidden phrases are the ones that make every follow-up email read identically. Requiring new useful information forces you to wait until you have a real reason to follow up rather than nagging. The “actual reason, not invented” line is the key constraint, because AI will fabricate a reason if you let it.

3. The reply to a long email

Why this works: Long inbound emails often hide three or four asks inside a wall of context. The summary-then-list output is the part that actually saves you time. The “defer with a date” constraint prevents the “I will get back to you on this” non-commitment that emails go to die in. Six-sentence ceiling forces priority.

4. The thank-you email after a meeting

Why this works: The “no thanks for your time” opener is the most counterintuitive constraint and the most useful. Specific opens get read; generic opens get archived. The “next step is a date or a question, not a vague hope” turns a polite email into a project-moving email.

5. The bad-news email

Why this works: Bad-news emails are where most people lean too hard on AI for cover, and the result is mush. The three structural choices are the real decisions. “Regret to inform you” is the exact phrase that turns a recoverable conversation into a defensive one. Three versions gives you a real comparison; the “what I do not want them to think” input is what stops AI from softening past the point of clarity.

6. The polite decline

Why this works: The biggest failure mode in decline emails is the over-explained no that invites a counter-offer. The “one reason at most, and only if it serves them” constraint forces you to keep the message clean. “Do not invent a referral” is the AI-specific failure: AI will fabricate “I would love to introduce you to X” when X is not someone you actually want to volunteer.

7. The intro email between two people

Why this works: Bad intro emails commit both parties to a meeting before they have agreed to it. The double-opt-in framing (give both an easy out) is the format that high-volume connectors use because it preserves the relationship even when one party is not interested. The “no I will let you two take it from here” close is the AI default; the real close is a specific question or a date.

8. The scheduling email

Why this works: Scheduling emails are where the AI default does the most damage to your time. Three sentences with clear windows and a single-click reply is the format that gets a meeting on the calendar in one round-trip. The “do not invent a URL” constraint is critical, because AI will hallucinate Calendly links if you do not block it.

What is the worst thing you can do with AI on email?

• Send without subject-line review. AI default subject lines pattern-match as spam. Rewrite every subject line in your own voice before sending.
• Use the same outreach template for 200 emails. Gmail and Outlook both fingerprint near-identical messages and bury them in promotional tabs. The whole point of prompts 1-2 is per-recipient tailoring.
• Let AI add an unrequested signature block. AI often inserts a “Best regards, [Your Name]” with quotes the recipient was never going to read. Use your real signature, not the generated one.
• Skip the pre-send AI-tells pass. Em-dashes in every paragraph, rhetorical-question closers, and “I hope this email finds you well” are the three patterns that signal “form letter” to a human reader in under 2 seconds.
• Ask AI to be funnier or warmer than you are. The tone mismatch is what makes recipients smell something off. Match your real voice. If you do not joke in real life, do not joke in the email.

What if you write 50 emails a week?

Save these prompts as Claude skills, one per situation. If you write a lot of the same kind of email (cold outreach for sales, replies to inbound, scheduling for client work), bundle the relevant subset as a plugin and invoke with one shortcut. The setup cost is 60 minutes; the time saved over 100 emails is 5-6 hours. After that, the system pays you every week.

What are common questions about AI for email?

Will Gmail flag my emails as spam if I use AI?

Not for using AI, but for sending AI patterns at scale. Gmail’s filters in 2026 look at the message signature (subject line phrasing, opener defaults, link density) and at sending behavior (volume, recipient diversity, reply rates). Per-message tailoring and the AI-tells edit are what keep you out of the promotions tab.

Should I disclose AI in business emails?

Only when asked, or when the email is high-stakes and trust is the deliverable. Most business email does not require disclosure. High-stakes contexts (a sympathy email, a hiring decision, a legal communication) do not benefit from AI involvement to begin with; write those yourself.

What is the right subject-line length?

4-7 words. Long enough to be specific, short enough to render fully on a phone screen. The line “Quick question about X” works far better than “Following up on our previous conversation about strategic alignment.” Prompt 1 enforces the limit.

Should I write the email or write the prompt?

For one-off emails to someone you know, write the email yourself; the AI setup time is not worth it. For email situations you face often (cold outreach, follow-ups, scheduling), write the prompt once and reuse it. The eight prompts above are the eight situations almost everyone faces enough to justify the prompt setup.

Where do these prompts come from?

They are the communication section of the larger AI Prompt Library. The Library has over 500 prompts across 33+ categories with three difficulty levels, including expanded communication coverage: sales sequences, customer-support replies, executive updates, and conflict resolution.

Sources to read next?

• Litmus state-of-email research
• HubSpot email marketing statistics
• Apollo outbound benchmarks
• Harvard Business Review on workplace communication
• Email marketing (Grokipedia)

You may also like

• Best Claude Prompts for Sales — cold outreach + objection scripts
• Best AI Prompts for Customer Support — the queue without burning out the team
• Best AI Prompts for Meeting Notes — decisions that survive the meeting
• How to Edit AI Out of Your Writing — strip AI patterns before you publish
• Prompt to Workflow: The AI Ladder — when to climb from prompt to workflow