I've sent thousands of cold emails over the past eight years. The first few hundred were terrible. The next few thousand were mediocre. And then AI showed up and somehow made most people's cold emails worse.

Here's what happened: everyone got access to Claude and ChatGPT, fed it a one-line prompt like "write a cold email to a SaaS founder," and blasted the result to 5,000 people. Reply rates cratered. Spam filters got smarter. And every decision-maker's inbox started reading like the same AI wrote every single message.

The problem isn't AI. The problem is lazy inputs produce lazy outputs. I've been using what I call the 3-input method with Claude for the past 14 months, and it consistently pulls 20-45% reply rates across different verticals. Not open rates. Reply rates. The difference is doing the work before you prompt.

This article breaks down exactly how it works, shows you side-by-side comparisons of generic vs. 3-input emails, introduces the personalization waterfall for prioritizing signals, and gives you five real templates from different industries. Everything's anonymized, but the structures are pulled from actual campaigns.

How to Write Cold Emails with Claude AI That Actually Get Replies

Why Most AI Cold Emails Fail

Let me show you what 90% of people do. They open Claude and type something like:

Write a cold email to the VP of Engineering at a mid-size SaaS company.
We offer web development services. Make it professional and persuasive.

Claude dutifully produces something like:

Subject: Elevating Your Web Development Strategy

Hi [Name],

I hope this message finds you well. My name is [Your Name] and I work with SaaS companies to build exceptional web experiences. We specialize in modern frontend development and have helped companies like yours improve performance and user engagement.

I'd love to schedule a 15-minute call to discuss how we might be able to help your team. Would next Tuesday or Thursday work?

This email is dead on arrival. Here's why:

  • "I hope this message finds you well" -- instant delete trigger
  • Zero specificity about the prospect or their company
  • No evidence you've done any research
  • The value prop is vague enough to apply to literally any company
  • It reads like AI because it is undirected AI

According to data from Instantly.ai's 2025 benchmark report, generic AI-generated cold emails average 2-3% reply rates. That's actually lower than well-crafted manual emails from 2020. We went backwards.

The fix isn't to stop using AI. It's to give AI something real to work with.

The 3-Input Method Explained

The core idea is simple: Claude (or any LLM) is a writing engine, not a research engine. When you make it do both jobs, it hallucinates research and produces generic copy. When you separate the jobs -- you research, Claude writes -- the output is dramatically better.

Three inputs. Every time.

Input 1: Company Research Brief

This is NOT "they are a B2B SaaS company in the healthcare space." That's a category, not research.

A real research brief looks like this:

They raised Series B ($28M) in October 2025. Their VP Engineering just posted on LinkedIn about migrating from a legacy PHP monolith to Next.js. Their careers page shows 3 open frontend roles and 1 DevOps position. Their marketing site Lighthouse score is 42. They recently launched a patient portal feature that's getting mixed reviews on G2 for slow load times.

See the difference? Every sentence is a hook Claude can use.

Input 2: ICP Pain Points for This Segment

Not generic pain points. Segment-specific ones you've validated through actual conversations with similar buyers.

Example for Webflow agencies:

Webflow agencies are losing $30K+ projects because Webflow can't handle auth, payments, or database operations. They're either turning down work or subcontracting it out at thin margins. They need a development partner who speaks Webflow's design language but can build the backend in Next.js or Astro.

Input 3: Tone Guide

This is the secret weapon most people skip. Without it, Claude defaults to corporate-speak.

Peer-to-peer tone. Not salesy. Short sentences. One question as the CTA -- no calendar links, no "would love to chat." Reference something specific from their website or LinkedIn. No "I hope this finds you well." No exclamation points. Write like a developer who noticed something interesting, not a sales rep working a list.

Side-by-Side: Generic vs. 3-Input Email

Let's use a real scenario. The prospect is the VP of Engineering at a healthcare SaaS company that just raised Series B.

The Generic Prompt

Write a cold email to a VP of Engineering at a healthcare SaaS company.
We build high-performance web applications. Keep it short.

The Generic Output

Subject: Helping Healthcare SaaS Companies Build Better

Hi Sarah,

I lead a web development team that specializes in healthcare SaaS applications. We've helped companies improve their web performance and ship features faster.

Would you be open to a quick call this week to explore potential synergies?

Dead. "Explore potential synergies" alone would get this archived.

The 3-Input Prompt

Write a cold email using these three inputs:

INPUT 1 -- COMPANY RESEARCH:
MedPortal raised $28M Series B in October 2025. VP Engineering (Sarah Chen)
posted on LinkedIn about migrating from PHP to Next.js. Careers page shows
3 open frontend roles. Their patient portal has a Lighthouse performance
score of 42. G2 reviews mention slow load times on the portal.

INPUT 2 -- SEGMENT PAIN POINTS:
Healthcare SaaS companies post-Series B are under pressure to ship faster
while meeting HIPAA compliance. Engineering teams are stretched thin from
the migration AND feature work simultaneously. They're hiring but can't
wait 3 months for new devs to ramp up.

INPUT 3 -- TONE GUIDE:
Peer-to-peer, developer to engineering leader. Short sentences. Reference
the Next.js migration specifically. One question as CTA. No "I hope this
finds you well." No corporate language. Sound like someone who's been
through this exact migration before.

The 3-Input Output

Subject: the PHP to Next.js migration

Hi Sarah,

Saw your post about the Next.js migration. We just finished a similar one for a HIPAA-compliant patient portal -- went from a Lighthouse score in the low 40s to 94 in about eight weeks.

With three frontend roles open and a Series B timeline to hit, I'm guessing the migration and new feature work are competing for the same people.

We've handled that exact overlap for two other healthtech teams this year. Worth comparing notes?

Night and day. Every sentence references something real. The CTA is low-pressure. It sounds like a peer, not a vendor.

How to Write Cold Emails with Claude AI That Actually Get Replies - architecture

The Personalization Waterfall

You won't always find a LinkedIn post about a tech migration. Some prospects are quieter online. The personalization waterfall tells you which signal to prioritize based on strength.

Priority Signal Type Example Why It Works
1 (Strongest) Hiring signals 3 open Next.js roles on careers page Shows active investment + specific tech decisions
2 Recent news Series B announcement, product launch Time-sensitive, shows you're paying attention
3 Tech stack change Migration posts, new tool adoption Technical credibility + pain point alignment
4 Client/performance data Lighthouse score of 42, slow G2 reviews Concrete problem you can solve
5 (Weakest) General positioning Company mission, industry vertical Better than nothing but low differentiation

Always use the highest signal available. If you find hiring signals, lead with those. If you can only find general positioning, you need to decide whether this prospect is even worth emailing -- because a Level 5 email will perform like a generic one.

The waterfall also compounds. The best emails use Level 1 + Level 2 together. Sarah's email above combines hiring signals (3 open frontend roles), recent news (Series B), and tech stack change (PHP to Next.js). That's three layers of proof that you actually did your homework.

Building Your Research Brief (Input 1)

This is where the real work happens. I spend 3-5 minutes per prospect. Here's my exact process:

  1. Careers page first. Open roles tell you what they're building, what stack they use, and where they're feeling pain. Three open frontend roles + one DevOps role tells a story.

  2. LinkedIn of the decision-maker. Not the company page -- the person. Recent posts, comments, shared articles. A VP Engineering sharing an article about monolith-to-microservice migrations is telling you what keeps them up at night.

  3. Run a Lighthouse audit. Takes 10 seconds. If their marketing site scores below 50, that's a concrete data point you can reference. We do a lot of Next.js development and Astro development work that starts with exactly this kind of finding.

  4. Check G2/Capterra reviews. Filter by recent negative reviews. Complaints about performance, slow dashboards, or clunky UX are gold.

  5. Crunchbase for funding/news. Recent rounds mean budget + pressure to execute.

The brief doesn't need to be long. Five sentences with five specific facts beats five paragraphs of fluff.

Crafting Segment-Specific Pain Points (Input 2)

This input comes from your own experience, not from researching the prospect. It's what you know about the category of buyer.

You build this once per segment and reuse it. Here are three examples:

For Webflow agencies losing enterprise deals:

These agencies cap out at $15-30K projects because Webflow can't handle authentication, payment processing, or database-driven features. They're turning down $50-100K builds or subcontracting the backend at thin margins. They need a headless CMS development partner who understands design-first workflows.

For post-Series B SaaS companies:

Engineering is stretched between maintaining the existing product and building what they promised investors. Hiring takes 3-6 months to ramp. They need senior-level capacity now without the overhead of full-time hires.

For manufacturing companies with legacy websites:

Their site was built in 2018 and hasn't been touched since. It doesn't work on mobile. Their competitors have configurators and real-time quoting tools. They're losing RFPs because purchasing managers Google them first.

Notice these aren't guesses. They're patterns from dozens of actual sales conversations. If you haven't talked to enough people in a segment to write this, you're not ready to email that segment.

The Tone Guide That Changes Everything (Input 3)

Claude's default tone is helpful, slightly formal, and recognizably AI. The tone guide overrides that.

My standard tone guide for cold emails:

Tone: Peer-to-peer. Developer writing to a technical leader.
Sentences: Short. Mostly under 15 words.
CTA: One question. No calendar links. No "would love to."
Banned phrases: "I hope this finds you well," "just reaching out,"
"potential synergies," "would love to chat," any exclamation points.
Specificity: Reference at least one fact from the research brief
in the first two sentences.
Length: Under 80 words total. Three paragraphs max.
Sign-off: Just a first name. No title, no company tagline.

The tone guide is where you inject your personality. If your brand voice is more casual, say so. If you're writing to C-suite at enterprise companies, adjust. The point is that you decide the voice, not Claude's defaults.

5 Real Cold Email Templates by Vertical

These are anonymized but structurally identical to emails that generated replies. Each shows the three inputs and the resulting email.

Template 1: Healthcare SaaS

Input 1: Patient engagement platform, Series A, hiring React developers, Lighthouse score 38, CEO posted about HIPAA-compliant infrastructure challenges.

Input 2: Healthcare startups post-funding struggle to find developers who understand both modern frontend frameworks and HIPAA compliance requirements. Most agencies don't know what a BAA is.

Input 3: Respectful but direct. Reference the HIPAA post. Short sentences. Ask if they've solved the problem yet.

Resulting email:

Subject: HIPAA + React -- solved or still painful?

Hi James,

Your post about HIPAA-compliant infrastructure hit close to home. We built the patient dashboard for [anonymized healthtech company] last year -- SOC 2 and HIPAA from day one, not bolted on after.

Noticed you're hiring React devs. If the pipeline's slow, we might be able to bridge the gap while you build the team.

Have you found a good setup for this yet?

-- Mike

Template 2: Manufacturing

Input 1: Industrial valve manufacturer, website last updated 2019, no mobile responsiveness, competitors have product configurators, 200+ SKUs listed as flat PDF catalogs.

Input 2: Manufacturing companies lose RFPs because engineers and purchasing managers can't find specs online. PDF catalogs don't get indexed by Google. Their competitors with modern sites are winning the search game.

Input 3: Straightforward, no jargon. Reference the PDF catalog specifically. Mention a competitor by category, not name. Question CTA.

Resulting email:

Subject: your valve catalog PDFs

Hi Tom,

I was looking at your product catalog and noticed it's all PDFs. That means Google can't index your individual valve specs -- engineers searching for specific models are finding your competitors instead.

We turned a similar PDF catalog (180+ SKUs) into a searchable, filterable product database for another industrial manufacturer. Their organic traffic went up 340% in five months.

Is the website something you're looking at changing this year?

-- Sarah

Template 3: Agency Partnership

Input 1: Webflow design agency, 12-person team, portfolio shows $10-20K branding projects, recently lost a pitch (founder mentioned it in a podcast) because the client needed user authentication.

Input 2: Webflow agencies hit a ceiling when projects require backend functionality. They either say no to the work or try to hack solutions with Memberstack/Outseta that don't scale.

Input 3: Collaborative, not competitive. Position as a partner, not a replacement. Reference the podcast. Casual tone.

Resulting email:

Subject: the auth problem from your podcast

Hi Jess,

Caught your episode on [podcast name] -- the part about losing the project because Webflow couldn't handle auth. That's painfully common.

We work with three Webflow agencies as their backend partner for exactly this. They handle design and content in Webflow, we build the auth, payments, and database layer in Next.js. The client gets one deliverable.

Is this something that keeps coming up for you?

-- Dan

Template 4: SaaS (B2B)

Input 1: Project management SaaS, 50-person company, marketing site runs on WordPress, page load time 6.2s, just launched a new pricing page, VP Marketing is posting about conversion rate optimization.

Input 2: B2B SaaS companies with WordPress marketing sites fight constant performance issues. Every second of load time costs conversions. They know it but the engineering team is focused on the product, not the marketing site.

Input 3: Data-driven. Lead with the speed number. Don't trash WordPress directly -- just state the performance fact. Question CTA.

Resulting email:

Subject: 6.2 seconds

Hi Rachel,

Your marketing site loads in 6.2 seconds on mobile. With the new pricing page you just launched, that load time is probably costing you signups -- every second above 3s drops conversion rates by about 7%.

We moved [anonymized SaaS company] from WordPress to Astro last quarter. Load time went from 5.8s to 1.1s. Their trial signups increased 22% with no other changes.

Is site performance something your team has bandwidth for right now?

-- Alex

Template 5: Higher Education

Input 1: Regional university, enrollment declining 8% YoY, website runs on a 2017 Drupal install, prospective student portal has 12-step application process, mobile bounce rate likely high (based on similar institutions).

Input 2: Universities with outdated web platforms are losing prospective students to competitors with modern, mobile-first application experiences. Gen Z expects instant, app-like interactions. A 12-step application form on a slow Drupal site is an enrollment killer.

Input 3: Respectful of higher ed culture. Don't be too casual. Reference the application process specifically. Offer a benchmark comparison. Question CTA.

Resulting email:

Subject: your application process on mobile

Hi Dr. Martinez,

I went through your prospective student application on my phone. It's 12 steps, and several of the form fields don't resize properly on mobile. For a generation that abandons checkout flows after two screens, that's a significant friction point.

We recently rebuilt the application portal for [anonymized university] -- reduced it to 3 screens with save-and-resume. Their completion rate increased 34%.

Would it be useful to see how your application experience compares to peer institutions?

-- Chris

Tools and Workflow for Scaling This

The 3-input method doesn't mean you can't scale. It means you scale the research, not just the sending.

Here's the stack I recommend in 2026:

Tool Purpose Cost (2026)
Apollo.io Lead sourcing + hiring signal data $49-99/mo
Clay Research automation + enrichment $149-349/mo
Claude API (Sonnet 4) Email generation via 3-input prompts ~$20/1M tokens
Instantly.ai Sending + warm-up + deliverability $37-97/mo
PageSpeed Insights Lighthouse scores for prospects Free
BuiltWith Tech stack identification $295/mo or free tier

Clay is the multiplier here. You can build workflows that automatically pull hiring signals from careers pages, run Lighthouse audits, and check BuiltWith data -- then format it all as Input 1 for Claude. That turns 5 minutes of research per prospect into 30 seconds of review.

The human stays in the loop for quality control. I review every email before it sends. At scale, that means I can personally approve 50-80 emails per hour, which is plenty for most outreach campaigns.

If you want help building a site that actually converts the traffic these emails generate, that's what we do at Social Animal -- check out our pricing or get in touch.

FAQ

Does the 3-input method work with ChatGPT or just Claude?

It works with any capable LLM. I prefer Claude (specifically Sonnet 4 as of mid-2025) because it follows tone instructions more consistently and produces less "AI-sounding" copy. GPT-4o works too, but you'll need to be more explicit in your tone guide to avoid its tendency toward enthusiastic corporate language.

How long does the research for Input 1 actually take?

About 3-5 minutes per prospect if you're doing it manually. With Clay or a similar enrichment tool automating the careers page scrape, Lighthouse audit, and BuiltWith lookup, it drops to about 30-60 seconds of review time. The LinkedIn check is still manual -- I haven't found a good way to automate reading someone's recent posts for context.

What reply rate should I expect with the 3-input method?

I've seen 20-45% across different verticals and list sizes. Healthcare and manufacturing tend to run higher (30-45%) because those buyers get fewer cold emails overall. SaaS and agency prospects are more saturated, so expect 15-25%. These numbers assume a clean list, proper warm-up, and good deliverability setup.

How many emails can I send per day using this approach?

With a single sending domain, 30-50 per day maximum. With 3-5 rotating domains (which Instantly handles well), you can push 100-200 per day while maintaining deliverability. The bottleneck is usually the review step, not the generation. I can review and approve about 60-80 emails per hour.

What if I can't find any strong personalization signals for a prospect?

Then don't email them yet. Seriously. A Level 5 personalization (general positioning only) performs almost identically to a fully generic email. Either wait for a signal to appear -- set up Google Alerts for the company, check back in a month -- or focus your time on prospects where you actually have something specific to say.

Should I include a calendar link in the CTA?

No. A calendar link in a first-touch cold email feels presumptuous. It assumes they've already decided to talk to you. A question CTA ("Is this something you're looking at?") gets significantly higher reply rates because it's a lower commitment. Save the calendar link for your follow-up after they reply positively.

How do I handle follow-ups with this method?

The first follow-up (3-4 days later) should add a new piece of value or a different angle from your research, not just "bumping this to the top of your inbox." The second follow-up (5-7 days after that) can be a simple two-line breakup email: "Looks like this isn't a priority right now. If the [specific thing from Input 1] becomes a pain point, I'm around." Three touches total. Then stop.

Is this method GDPR compliant for European prospects?

The 3-input method itself is just a prompting framework -- GDPR compliance depends on how you source your contact data and whether you have a legitimate interest basis for the outreach. Using publicly available business contact information for B2B outreach is generally acceptable under legitimate interest, but you need to include an opt-out mechanism and honor removal requests immediately. If you're targeting EU prospects at any volume, get a lawyer's opinion for your specific situation.