Claude API Email Generation Templates

To generate personalized emails with the Claude API, pass your recipient data as variables into a structured system prompt that defines tone, intent, and constraints, then render each message in the user turn. Claude returns ready-to-send copy — no post-processing required. For cold outreach, use claude-haiku-4-5 at $0.0002 per email; for high-value sequences where tone precision matters, switch to claude-sonnet-4-6 at $0.0030. A single Batch API job can generate 1,000 personalized emails overnight at half the synchronous cost.

Prompt Patterns for Cold, Warm, and Transactional Emails

Cold Email

Cold email prompts need a firm word-count ceiling, a clear value hook, and an explicit instruction to avoid spam triggers (more on that in Deliverability Considerations).

import anthropic

client = anthropic.Anthropic()

COLD_EMAIL_SYSTEM = """You are an expert B2B copywriter.
Write a cold outreach email following these rules exactly:
- Subject line: max 50 characters, no all-caps, no exclamation marks
- Body: 80-120 words, plain conversational tone
- One specific pain point tied to the recipient's industry
- One concrete offer or question as the call to action
- No generic openers ("I hope this finds you well", "My name is...")
- No spam trigger words: free, guaranteed, act now, limited time, winner
Return format:
Subject: [subject line]

[email body]"""

def generate_cold_email(recipient: dict) -> str:
    prompt = (
        f"Recipient: {recipient['name']}, {recipient['title']} at {recipient['company']}.\n"
        f"Industry: {recipient['industry']}.\n"
        f"Our product: {recipient['our_product']}.\n"
        f"Known pain point: {recipient['pain_point']}."
    )
    response = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=400,
        system=COLD_EMAIL_SYSTEM,
        messages=[{"role": "user", "content": prompt}],
    )
    return response.content[0].text

Warm Email (Existing Lead)

Warm emails can reference past interactions. Include a last_touchpoint variable so Claude anchors the opening to something real.

WARM_EMAIL_SYSTEM = """You are a helpful sales advisor writing a follow-up email.
Rules:
- Reference the last touchpoint naturally in the first sentence
- Body: 60-90 words
- Offer one concrete next step (demo, resource, or question)
- Tone: friendly but professional, no pressure language
Return Subject + body only."""

def generate_warm_email(lead: dict) -> str:
    prompt = (
        f"Name: {lead['name']}. Last touchpoint: {lead['last_touchpoint']}.\n"
        f"Their interest: {lead['interest']}. Next step to offer: {lead['next_step']}."
    )
    response = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=350,
        system=WARM_EMAIL_SYSTEM,
        messages=[{"role": "user", "content": prompt}],
    )
    return response.content[0].text

Transactional Email

Transactional emails (order confirmations, password resets, receipts) are low-creativity, high-accuracy tasks — ideal for Haiku with a strict JSON output contract.

TRANSACTIONAL_SYSTEM = """You write concise transactional emails.
Return a JSON object: {"subject": "...", "body": "..."}
Rules:
- Subject: factual, max 45 chars
- Body: under 60 words, no marketing language
- Include all provided data fields verbatim (order ID, amounts, dates)
- Plain text only — no markdown, no HTML"""

def generate_order_confirmation(order: dict) -> dict:
    import json
    prompt = json.dumps(order)
    response = client.messages.create(
        model="claude-haiku-4-5",
        max_tokens=300,
        system=TRANSACTIONAL_SYSTEM,
        messages=[{"role": "user", "content": prompt}],
    )
    return json.loads(response.content[0].text)

Personalization Variables

Effective personalization goes beyond {first_name}. The variables below consistently lift open and reply rates:

Pass all available variables into the prompt. Claude uses them selectively — you do not need to engineer which ones appear in the output. For data-sparse recipients (missing pain_point, for example), add a fallback instruction: "If a field is empty, omit that element rather than inventing a placeholder.".

A/B Testing Email Variants

Generate two variants per recipient in a single call by asking for a JSON array. Run the split in your CRM or email tool.

AB_SYSTEM = """You are a conversion copywriter.
Generate TWO cold email variants for A/B testing.
Variant A: leads with a pain-point question.
Variant B: leads with a bold proof statement (stat or outcome).
Return JSON: [{"variant": "A", "subject": "...", "body": "..."}, {"variant": "B", ...}]
Each body: 80-110 words. Same recipient data, different angle."""

def generate_ab_variants(recipient: dict) -> list[dict]:
    import json
    response = client.messages.create(
        model="claude-sonnet-4-6",  # Sonnet for sharper variant differentiation
        max_tokens=700,
        system=AB_SYSTEM,
        messages=[{"role": "user", "content": json.dumps(recipient)}],
    )
    return json.loads(response.content[0].text)

Use Sonnet for A/B variant generation — the quality difference between Haiku and Sonnet is most visible when you need two meaningfully distinct angles rather than surface-level rewrites. See Claude Haiku vs Sonnet vs Opus — Which Model for the full decision framework.

Batch Generation with Batch API

The Batch API runs asynchronous jobs at a 50% discount off synchronous pricing. For 1,000-email campaigns, this cuts cost from $0.20 (Haiku sync) to $0.10 per batch run.

import anthropic
import json

client = anthropic.Anthropic()

def build_batch_requests(recipients: list[dict]) -> list[dict]:
    """Build a Batch API request list from a recipient list."""
    requests = []
    for i, r in enumerate(recipients):
        prompt = (
            f"Recipient: {r['name']}, {r['title']} at {r['company']}.\n"
            f"Industry: {r['industry']}. Pain point: {r['pain_point']}."
        )
        requests.append({
            "custom_id": f"email_{i}_{r['email']}",
            "params": {
                "model": "claude-haiku-4-5",
                "max_tokens": 400,
                "system": COLD_EMAIL_SYSTEM,
                "messages": [{"role": "user", "content": prompt}],
            },
        })
    return requests


def submit_email_batch(recipients: list[dict]) -> str:
    """Submit a batch job and return the batch ID."""
    requests = build_batch_requests(recipients)
    batch = client.messages.batches.create(requests=requests)
    print(f"Batch submitted: {batch.id} — {len(requests)} emails queued")
    return batch.id


def collect_batch_results(batch_id: str) -> dict[str, str]:
    """Poll until complete, then return {custom_id: email_text}."""
    import time

    while True:
        batch = client.messages.batches.retrieve(batch_id)
        if batch.processing_status == "ended":
            break
        print(f"Status: {batch.processing_status} — waiting 30s")
        time.sleep(30)

    results = {}
    for result in client.messages.batches.results(batch_id):
        if result.result.type == "succeeded":
            results[result.custom_id] = result.result.message.content[0].text
        else:
            results[result.custom_id] = None  # log failures separately
    return results


# --- Usage ---
# recipients = load_from_crm(limit=1000)
# batch_id = submit_email_batch(recipients)
# emails = collect_batch_results(batch_id)

For concurrent synchronous sends (when you need results in under a minute), see Claude API Concurrent Requests for an asyncio-based approach.

Agent SDK Cookbook — Email Automation Recipes

Agent SDK Cookbook — $49

The cookbook includes complete email agent blueprints: multi-step outreach sequences, reply-detection loops, CRM sync patterns, and Haiku/Sonnet routing logic for cost-optimized campaign generation. All recipes ship with prompt caching enabled and Batch API integration out of the box.

→ Get the Agent SDK Cookbook — $49

Instant download. 30-day money-back guarantee.

Deliverability Considerations: Avoiding Spammy Patterns

AI-generated email fails deliverability for two distinct reasons: content triggers (spam-filter keywords) and structural red flags (identical copy sent to many recipients). Address both in the prompt and in your sending infrastructure.

Prompt-level controls:

DELIVERABILITY_RULES = """
Banned words and phrases (never use):
- free, guaranteed, no risk, act now, limited time offer, click here
- winner, congratulations, you have been selected
- $$, 100%, earn money, make money fast
- Excessive punctuation: !!!, ???

Structural rules:
- Vary sentence length — avoid uniform 15-word sentences throughout
- No ALL CAPS words except established acronyms (API, SaaS, CRM)
- Include one genuine question to invite a reply
- Sign off with a real name and role, not a generic "The Team"
"""

Infrastructure controls (outside Claude):

• Send from a warmed domain (at least 30 days old with gradual volume ramp)
• Limit daily send volume to under 200 per domain until reputation is established
• Use a unique custom_id per recipient and deduplicate before sending
• Add plain-text MIME part alongside HTML — spam filters penalize HTML-only emails
• Honor unsubscribe immediately; bounce rate above 2% triggers ISP throttling

Even with perfect copy, high send volume from a cold domain will land in spam. The Claude prompt controls content quality; your ESP configuration controls inbox placement.

Template Versioning Pattern

As your prompts evolve, version them explicitly. This lets you A/B test prompt versions across campaigns and roll back if quality drops.

from dataclasses import dataclass
from datetime import date

@dataclass
class EmailTemplate:
    version: str
    model: str
    system: str
    max_tokens: int
    created: str

TEMPLATES: dict[str, EmailTemplate] = {
    "cold_v1": EmailTemplate(
        version="cold_v1",
        model="claude-haiku-4-5",
        system=COLD_EMAIL_SYSTEM,
        max_tokens=400,
        created="2026-04-01",
    ),
    "cold_v2": EmailTemplate(
        version="cold_v2",
        model="claude-haiku-4-5",
        system=COLD_EMAIL_SYSTEM + "\n- Open with a specific industry stat if possible.",
        max_tokens=450,
        created="2026-04-30",
    ),
}


def generate_with_template(template_key: str, recipient: dict) -> dict:
    tpl = TEMPLATES[template_key]
    prompt = (
        f"Recipient: {recipient['name']}, {recipient['title']} at {recipient['company']}.\n"
        f"Industry: {recipient['industry']}. Pain point: {recipient['pain_point']}."
    )
    response = client.messages.create(
        model=tpl.model,
        max_tokens=tpl.max_tokens,
        system=tpl.system,
        messages=[{"role": "user", "content": prompt}],
    )
    return {
        "template_version": tpl.version,
        "model": tpl.model,
        "email": response.content[0].text,
        "input_tokens": response.usage.input_tokens,
        "output_tokens": response.usage.output_tokens,
    }

Log template_version and token counts per send. This data feeds directly into the break-even analysis below.

Cost Analysis: Per-Email Cost and Break-Even

Per-email cost (April 2026 list pricing)

Typical cold email: ~250 input tokens (system + recipient data) + ~150 output tokens.

For a 1,000-email campaign: Haiku sync ≈ $1.00, Haiku batch ≈ $0.40, Sonnet sync ≈ $3.00.

Break-even vs. human copywriter

A mid-level freelance copywriter charges $5–$15 per personalized cold email. At $0.0008 per email with Haiku, Claude reaches break-even in fewer than 1 email — there is no cross-over point; Claude is cheaper at every volume. The real comparison is Claude + human review vs. fully human:

For prompt caching strategies that reduce input token costs by up to 90% on repeated system prompts, see Claude API Cost & Prompt Caching Break-Even.

Frequently Asked Questions

How do I generate personalized emails at scale without hitting rate limits?

Use the Batch API for non-time-sensitive campaigns (results available within minutes to hours). For real-time personalization at scale, run concurrent async requests with asyncio — see Claude API Concurrent Requests. Haiku's rate limits are significantly higher than Sonnet's, making it the right choice for bulk sends.

Will AI-generated emails get flagged as spam?

Claude-generated copy avoids spam triggers when you include explicit banned-word rules in the system prompt (see the Deliverability section above). The bigger deliverability risk is not content quality but sending infrastructure: domain age, DKIM/DMARC setup, and gradual volume ramp. Spam filters evaluate the envelope (sender reputation) before they evaluate the body. Clean copy from a cold domain will still land in spam.

What is the best model for cold email — Haiku or Sonnet?

Haiku is the right default for cold email at scale. The quality gap between Haiku and Sonnet is small for structured short-copy tasks like 100-word emails, and Haiku is ~3.75x cheaper per email. Use Sonnet when you need sharper A/B variant differentiation, nuanced executive-level tone, or when you are writing fewer than 100 emails and cost is not a constraint. See Claude Haiku vs Sonnet vs Opus — Which Model.

Can Claude generate full multi-step email sequences?

Yes. Structure the sequence as a loop: pass the previous email's text as context in the next call so Claude maintains tone and avoids repeating the same hooks. For a 5-email drip sequence, generate all five in a single session with a rolling context window rather than five independent calls — this produces more coherent sequences and reduces redundancy.

How do I version and roll back email prompt templates?

Use the EmailTemplate dataclass pattern above. Store templates in a config file or database keyed by version string. Log the template_version and token counts for every send. When a campaign underperforms, compare reply rates by version to identify which prompt change caused the regression. Never overwrite a live template in place — always create a new version key.

Agent SDK Cookbook — Full Email Automation Agent

Agent SDK Cookbook — $49

Go beyond single-email generation: the cookbook's email chapter covers a full agentic pipeline — CRM data ingestion, multi-step sequence generation, reply detection, and automated follow-up scheduling. Built on the Anthropic Agent SDK with Batch API integration, prompt caching, and template versioning baked in.

→ Get the Agent SDK Cookbook — $49

Instant download. 30-day money-back guarantee.

Sources

1. Anthropic — Claude model pricing — April 2026
2. Anthropic — Message Batches API — April 2026
3. Anthropic — Rate limits — April 2026