Hey folks, been experimenting with Claude 3 Opus for automating some cold outreach personalization at my startup. We're sending a few hundred emails a week and wanted to move beyond basic mail-merge.
The initial results are pretty mind-blowing for research-heavy personalization. Feed it a prospect's LinkedIn bio, recent company news, and a product one-pager, and it writes something genuinely relevant. Way better than the "I see you work at [Company]" stuff.
But scaling it up? Ran into some real hiccarts.
First, the context window is a blessing and a curse. You can dump a ton of research, but the latency for each email generation adds up. We built a simple pipeline using the API, but the cost per email (with Opus) gets noticeable when you're doing volume. We're testing Sonnet now for speed/cost.
The bigger pitfall is the "politeness drift." Claude *really* wants to be helpful and ethical, which is great, but it sometimes over-corrects. It would add lengthy, softening disclaimers like "I know you're incredibly busy" or "If this isn't relevant, please disregard," which watered down the message. We had to get super specific in the system prompt to clamp down on that.
Here's a snippet of the prompt structure we landed on:
```markdown
Role: Senior Sales Development Representative.
Task: Write a concise, value-driven cold email.
Inputs:
- Prospect Name: {name}
- Prospect Role & Company: {role} at {company}
- Recent Trigger Event: {event}
- Our Value Prop: {value}
Rules:
- Maximum 4 sentences.
- Lead with the trigger event (their news).
- Immediately connect to our value prop.
- End with a clear, low-barrier call to action.
- Do not use apologies or "just checking in" phrases.
- Tone: confident, professional, direct.
```
Anyone else trying this? Curious if you've found tricks to keep the tone sharp without needing a paragraph of prompt constraints. Also, how are you handling the research data ingestion—chunking it yourself before sending to the API, or using Claude's full context? The token cost on that is my next thing to optimize.
The politeness drift is a real issue. It stems from the model's training data and safety guardrails. You're right to combat it with a strict system prompt, but that often isn't enough.
You need to provide negative examples in your few-shot prompts. Show it exactly what you *don't* want. Include an example of an overly soft, disclaimer-filled email and label it as "unacceptable style - too indirect."
Also, benchmark the cost per qualified reply, not per email. If Opus gets a 2% reply rate and Sonnet gets 1.5%, the extra cost might be justified. Most people just look at the raw API cost and miss the conversion math.
page 18 is where the traps are