How We Use Claude to Write Truly Personalized Cold Emails

Open any AI SDR product page and you will see the same claim: "AI-powered personalization." Click deeper and you discover what that actually means. The AI inserts the prospect's first name and company into a pre-written template. Dear {first_name}, I noticed {company} is growing fast. That is not personalization. That is mail merge with a marketing budget.

We built GetSalesClaw to do something fundamentally different. Every email that GetSalesClaw sends is written from scratch by Claude, Anthropic's large language model. No templates. No placeholder variables. No pre-written paragraphs with swappable fields. Each email is a unique piece of writing, generated after the AI has analyzed the prospect's company, role, industry, and context.

This article explains exactly how that works, from the models we chose to the pipeline that runs before a single word is written.

Why We Chose Claude

When we started building GetSalesClaw, we evaluated every major language model on the market: OpenAI's GPT-4, Google's Gemini, Mistral, Llama, and Anthropic's Claude. The decision came down to three factors that matter specifically for cold email generation.

Instruction-following. Cold email generation requires the AI to follow detailed constraints simultaneously: match a specific tone, stay under a word count, reference particular company details without fabricating others, write in a specific language, and avoid certain phrases. Claude consistently followed multi-layered instructions more reliably than alternatives we tested. When you tell Claude to write a 4-sentence intro email in French that references a company's recent product launch without being pushy, it does exactly that. Other models would drift: ignoring the word count, switching to English mid-paragraph, or defaulting to aggressive sales language despite explicit instructions not to.

Natural tone. Most AI-generated cold emails sound like AI-generated cold emails. There is a recognizable pattern: overly enthusiastic opening, generic value proposition, forced call-to-action. Claude produces writing that reads like it came from a thoughtful person who actually looked at the prospect's company before writing. That distinction matters because recipients can tell. They have seen hundreds of AI emails by now. An email that sounds genuinely human gets read. An email that sounds like ChatGPT gets archived.

Multilingual quality. GetSalesClaw operates primarily in European markets. Our users write cold emails in French, English, German, Spanish, Italian, Dutch, and Portuguese. Many AI models handle English well but produce stilted, unnatural output in other languages. Claude's multilingual capabilities are strong across all seven languages we support. A French startup founder sending emails to German prospects gets output that reads naturally in German, not like it was translated from English. This is not a nice-to-have. In European B2B sales, sending a poorly written email in someone's native language is worse than sending a good one in English.

Transparency. We also made a deliberate decision to be transparent about which AI powers GetSalesClaw. Many competitors say "proprietary AI" or "our advanced AI engine," which in practice means they are calling the same APIs everyone else uses but do not want you to know that. We use Claude by Anthropic. We say so publicly. If Anthropic improves Claude, GetSalesClaw gets better. If a customer asks what generates their emails, we give a straight answer. There is no competitive advantage in obscuring your AI provider. The advantage is in how you use it.

The Two-Pass Scoring System

Before GetSalesClaw writes a single email, it scores every lead through a two-pass system. This is critical because AI-generated emails cost money (both in API calls and in your domain reputation), and sending a beautifully personalized email to the wrong person is still a waste.

Pass 1: Fast filter with Claude Haiku

The first pass uses Claude Haiku, Anthropic's fastest and most cost-efficient model. Haiku acts as a quick filter. It receives the lead's basic information (company name, industry, size, role) and your Ideal Customer Profile, and makes a fast yes/no/maybe decision.

This pass eliminates obvious non-matches: companies in the wrong industry, organizations that are too small or too large, contacts who are not decision-makers, businesses in geographies you do not serve. The goal is speed and cost efficiency. Haiku processes each lead in under a second for a fraction of a cent.

Roughly 40-60% of raw leads get filtered out at this stage. That is 40-60% of leads that never reach the expensive scoring model, never get an email written, and never waste your sending reputation. The ROI of this single filtering step pays for itself many times over.

Pass 2: Deep analysis with Claude Sonnet

Leads that pass the Haiku filter move to Claude Sonnet, Anthropic's mid-tier model that balances intelligence with cost. Sonnet performs a detailed analysis and assigns a score from 0 to 100.

This is not a simple keyword match. Sonnet analyzes multiple dimensions:

• Company profile. What does the company do? What industry are they in? How large are they? What stage are they at?
• Technology signals. What tools and platforms does the company use? Do they indicate a need for your product?
• Industry position. Is the company growing, stable, or contracting? Are they in a market segment that aligns with your solution?
• Contact relevance. Is this person a decision-maker? Do they have budget authority? Is their role directly related to the problem you solve?
• Timing signals. Has the company recently raised funding, launched a product, expanded into new markets, or made a hire that suggests they need what you sell?

The output is a numerical score with a written explanation. A score of 82 means something specific: the model explains why this lead is a strong match. A score of 35 comes with reasons: "Company matches on industry but is too early-stage and the contact is in a non-decision-making role." This transparency matters because it lets users understand and refine their targeting over time.

Only leads scoring above a configurable threshold (default: 60) proceed to email generation. The rest are logged but not acted upon. You can review them later if you want to adjust your criteria.

How the Email Generation Works

This is where the real differentiation happens. Most AI SDR tools maintain a library of email templates and use AI to fill in the blanks. GetSalesClaw does not use templates at all. Every email is generated from a detailed context package that Claude receives fresh each time.

Context injection

When GetSalesClaw generates an email for a lead, Claude receives a carefully assembled context package:

• Company profile. Everything the scoring system learned about the prospect's company: what they do, their size, their industry, recent news or signals.
• Contact profile. The prospect's role, seniority, and what that role typically cares about.
• Your pitch (USER.md). A document you write once that describes your product, your ideal customer, your value proposition, and any specific angles you want emphasized. This is your voice, your positioning, your strategy.
• Writing guidelines (SOUL.md). A system-level document that defines tone, structure, length constraints, and anti-patterns. This is where we encode rules like "never use the phrase 'I hope this email finds you well'" and "keep the email under 120 words."
• Language and cultural context. The target language, and cultural norms for business communication in the prospect's region. A cold email to a German CTO reads differently from one to a French marketing director, even if the product pitch is the same.

No templates, no placeholders

Claude receives all of this context and writes the email from scratch. There is no template with blanks to fill. There is no library of opening lines to choose from. The model generates a complete email that is unique to this specific prospect at this specific company.

This means that if you have two leads at competing companies in the same industry, they will receive different emails. The value proposition might be similar, but the framing, the opening line, the specific reference points, and the call-to-action will differ based on what the AI learned about each company during scoring.

Three-email sequences

GetSalesClaw generates a three-email sequence for each qualified lead:

1. Intro email (Day 0). The first touch. Short, specific, focused on one clear value proposition relevant to the prospect's situation. No generic "I'd love to chat" closings. The CTA is concrete: a specific question, a relevant link, or a clear next step.
2. Follow-up (Day 3). A different angle on the same value proposition. Not a "just following up" email (those get deleted immediately). This email adds new information, a different perspective, or addresses a likely objection. It references the first email without repeating it.
3. Last chance (Day 7). A brief, direct email. Acknowledges that the prospect is busy, offers one final clear reason to respond, and makes it easy to say no. Respectful closing that does not burn the relationship for future outreach.

All three emails are generated together so they form a coherent sequence. The follow-up references the intro naturally. The last-chance email does not repeat points already made. This coherence is something template-based systems struggle with because each template is typically written independently.

Tone and cultural adaptation

The AI adapts its writing style based on the prospect's language and cultural context. French business emails tend to be more formal in the opening and more relationship-oriented. German business communication values directness and specific technical details. American-style emails with bold claims and casual tone can feel abrasive in European markets.

GetSalesClaw handles this automatically. When the system detects that a prospect is based in Germany and their business communication is in German, Claude generates an email that follows German business email conventions. The user does not need to maintain separate templates for each market. The AI adapts based on context.

The Human-in-the-Loop Safeguard

Here is where we diverge from every AI SDR that is racing toward full autopilot. GetSalesClaw requires human approval before sending any email.

Every generated email sequence is sent to the user on Telegram or Slack for review. One tap to approve. One tap to reject. If you reject, you can optionally tell the system why, and it learns from that feedback for future emails.

Why do we insist on this? Because no language model is perfect. Claude is excellent. It is the best model we have tested for this task. But it occasionally makes mistakes:

• It might reference a company detail that is outdated or incorrect.
• It might misjudge the tone for a specific prospect or situation.
• It might produce an email that is technically fine but does not match your current strategic priority.

These edge cases are rare (we estimate less than 5% of generated emails need rejection), but when they happen, catching them before send is the difference between building a relationship and burning one. At B2B deal sizes, a single bad email can cost you a prospect worth $10,000 or more. Five minutes of daily review eliminates that risk entirely.

We wrote a full article on this philosophy: Why We Refuse Full Autopilot in AI Sales.

The Results

The combination of deep scoring and from-scratch email generation produces measurably different results compared to template-based approaches.

Higher reply rates. Prospects respond more frequently to emails that reference their specific situation. A generic "I noticed your company is growing" gets ignored. A specific "I saw that [Company] just expanded into the Benelux market and is hiring three account executives there" gets read and often gets a reply, even if it is a polite "not right now."

Lower unsubscribe rates. When emails feel personal and relevant, recipients are less likely to hit unsubscribe or mark as spam. This protects your domain reputation long-term, which is critical for sustained outbound success.

Better conversations. When a prospect does reply, the conversation starts at a higher level. They are not asking "who are you and why are you emailing me?" They are responding to a specific point you made about their business. That is a fundamentally different starting position for a sales conversation.

Cost efficiency. Despite using premium AI models for generation, the average cost per email is approximately $0.01. The two-pass scoring system ensures that expensive email generation only happens for leads that are actually worth contacting. You are not paying to write emails to unqualified leads.

What This Means for You

If you are evaluating AI SDR tools, ask one question: does the AI write each email from scratch, or does it fill in templates? The answer determines whether your prospects receive something that feels like a thoughtful message from a real person or another piece of automated noise in their inbox.

Template-based tools will always produce emails that feel templated. Prospects have pattern-matching instincts. They can tell when an email follows a formula, even if the variables are filled in correctly. "Hi [Name], I noticed [Company] is [doing something]. We help companies like yours [value prop]. Would you be open to [CTA]?" That structure is recognizable after the third time you receive it.

From-scratch generation produces emails that do not follow a predictable pattern. Each one reads differently because each one is different. That unpredictability is what makes them feel human, and what makes prospects actually read them.

We are transparent about how GetSalesClaw works because we believe the approach speaks for itself. Claude writes your emails. A two-pass scoring system ensures they go to the right people. You review every email before it sends. And the whole thing costs about two cents per lead.