TL;DR
The average cold email reply rate fell to 3.43% in 2026, down from 8.5% in 2019 (Mailforge benchmark). Mass templates are dead. But top-quartile senders still hit 15-25% reply rates by doing one thing AI is genuinely good at: writing a specific note for each recipient, fast. This guide gives you the exact prompt, the per-channel length rules, and the follow-up cadence that captures 93% of replies by day 10.
The model does the drafting. You supply the signal it cannot find: the one public artifact (a post, a talk, an old role) that proves you actually looked.
The task
You want a 20-minute call with someone at a target company. No warm intro. You have to send a cold note that doesn’t land in their LinkedIn-spam mental bucket.
Personalization is the only lever left. Hunter.io’s analysis of 11 million emails found that depth of personalization (not merge tags) drives roughly 52% higher reply rates, and small, tightly-targeted campaigns outperform broad blasts by about 2.76x. AI lets you keep that depth while sending to 10-20 people instead of 1.
When this is the right job for AI
- You can hand AI a real reason you’re reaching out to this specific person (their post, their team, their old role).
- You have 5+ targets, so per-message effort actually matters.
- You will paste back signal data from the recipient’s public profile — AI cannot find this for you, and inventing it is how you get caught.
If you only have one target and unlimited time, write it yourself. AI’s edge here is volume with retained specificity, not a single perfect note.
What to feed the AI
- Recipient: name, current role, company, and 1 specific public artifact (talk, post, project, GitHub repo).
- Sender: your role, what you’re actively exploring, the 1-line reason you’re writing this person.
- Ask: a 20-minute call about X, with a fallback (2 async questions).
- Channel + limit: see the table below — the right length depends entirely on the channel.
Channel length rules (verified June 2026)
| Channel | Hard cap | What actually performs | Source |
|---|---|---|---|
| LinkedIn connection-request note | 300 chars (Premium) / 200 (Free) | 120-180 chars beats notes that fill the full 300 | Evaboot |
| LinkedIn DM (already connected) | 8,000 chars | 50-90 words; the cap is irrelevant | Evaboot |
| LinkedIn InMail | 2,000 chars body, 200 subject | Under 90 words | Evaboot |
| Cold email | none | Under 80 words, one CTA | Instantly 2026 report |
The common mistake is treating a LinkedIn DM like a connection note and cramming everything into 300 characters, or treating a connection note like an email and writing 250 words. Match the format.
Copy-ready prompt
Paste this into ChatGPT, Claude, or Gemini. As of June 2026, any flagship model (GPT-5.5, Claude Sonnet 4.6, or Gemini 3) handles this at the free tier; the $20/mo paid tiers help mainly when you batch 15+ at once and want the larger context window to hold all your targets.
You are writing a cold LinkedIn connection-request note to a hiring manager.
Recipient: Sam Otieno, VP Platform at Acme Cloud. Recent public artifact:
a Substack post on building the platform team's on-call culture.
Sender: I'm a senior platform engineer at GammaInc, exploring my next role
specifically in dev-tools companies that take on-call seriously.
Ask: 20-minute call to learn how Acme thinks about platform-team org design.
Fallback: 2 async questions.
Channel: LinkedIn connection-request note, hard cap 300 characters,
target 160 characters.
Rules:
- First line is the specific reason (cite the artifact).
- Do NOT write "I admire your work" or "love what you're building".
- Be explicit about what I want and what I am offering (questions, not a free favor).
- One line of credibility, not a resume dump.
- End with a one-sentence ask. No double ask.
- Count the characters and report the total.Swap the channel line to cold email, target 70 words or LinkedIn DM, target 60 words and the model adapts the length automatically — that’s why the limit lives in the prompt, not in your head.
Sample output (connection note, ~190 chars)
“Sam — your Substack on the 2-day on-call rotation hit home. I shipped a similar rotation at GammaInc (pager load -40%) and I’m filtering my next platform role on on-call seriousness. Open to a 20-min call, or 2 async questions if easier?”
Sample output (email version, ~75 words)
Subject: 2-day on-call rotation — quick question
Sam — your Substack on Acme’s on-call culture was the first piece I’ve seen treat on-call as an org-design question. I shipped a similar 2-day rotation at GammaInc (pager load dropped ~40%, on-call satisfaction up) and I’m exploring my next role with that as a filter.
Open to a 20-minute call in the next two weeks? Happy to send 2 questions in writing instead if that’s easier.
— [Name], Senior Platform Engineer, GammaInc
How to refine the output
- Sounds salesy → add: “No ‘I admire’, no ‘love what you’re building’, no ‘would love to chat’. Be specific or be silent.”
- Too long → tighten the limit and add “count the words and report the total.” Models overshoot; make them self-audit.
- Generic credibility → require one number from your own work (the
-40%above does more than any adjective). - Double ask → strict rule: “One ask only. The fallback is offered, not a second ask.”
- Reads like AI → ban the tells directly: “No em-dashes in every sentence, no ‘thrilled’, no ‘reach out’, no ‘circle back’.”
Follow-up cadence (this is where most replies actually come from)
The first message earns only about 58% of total replies; the rest come from follow-ups (Instantly). For sales, a 3-7-7 cadence (day 0, day 3, day 10) captures ~93% of replies by day 10. For job-search outreach, dial that down — you’re a human asking a favor, not a sequence:
- One follow-up, 7-10 days after the first, shorter than the original.
- Add something new (a question, a relevant link), never “just checking in.”
- Stop after the second message. Two no-replies is a no.
Send Tuesday-Thursday, 8-11 AM or 2-4 PM in the recipient’s timezone — the window with the strongest engagement across 2025-2026 platform data.
Common mistakes
- Same template to 50 people. Recipients compare notes; four friends spotting the same DM in 24 hours is how your name gets a reputation.
- Burying the ask. The first 1.5 lines must say why you’re writing.
- Asking for an hour. 20 minutes — or async questions — is the right shape.
- Faking a connection (“we both went to…”). If it’s thin, drop it. A weak connection reads worse than no connection.
- Letting AI invent the artifact. If the model doesn’t have a real post or project from you, it will fabricate one. Never send a reference you didn’t verify.
FAQ
- What’s a realistic reply rate? Top-quartile personalized job-search outreach lands 15-25%, versus the 3.43% market average for generic cold email (June 2026). Below 10% means your specificity is broken, not your volume.
- LinkedIn or email? Whichever is easier to find a real address/handle for. A LinkedIn connection note (300-char cap) signals “I saw your work”; email gives you room for one concrete number. Don’t write 250 words into a 300-character box.
- Does it cost money? No. The free tiers of ChatGPT, Claude, and Gemini all handle this. The $20/mo plans (ChatGPT Plus $20, Claude Pro $20, Google AI Pro $19.99) only help when you batch 15+ targets and want a bigger context window — Gemini’s ~1M-token window holds a whole spreadsheet of prospects at once.
- How many follow-ups? One, 7-10 days later, shorter than the first. Stop after the second. Job-search outreach is not a sales sequence.
- What if they reply but say no? Reply once, thank them, and paste the 2 async questions — many people who decline a call will still answer a tight written question.
Related
- AI networking follow-up
- AI recruiter reply prep
- AI company / role research
- AI thank-you email after interview
- Networking Outreach Prompts for Cold DMs and Emails
- AI Case Interview Prep Sparring Partner
Tags: #AI writing #Job search #Networking #Email writing #LinkedIn