After a few hundred sends, I figured out something embarrassing: when the Cold Email Written From a Prospect's Actual Website prompt gave me a weak email, it was almost never the model's fault. It was mine. I had pasted in the wrong part of the website. Garbage in, polite garbage out. So this is the practical guide I wish I'd had on day one, written so you skip the two weeks of bad sends I went through.
The Wrong Paragraph vs the Right Paragraph
Most people paste the hero headline. The hero headline is the worst possible input, because it is marketing-polished to within an inch of its life and tells you nothing real. 'We help teams do their best work' generates an observation as bland as the headline. What you actually want is the boring, revealing copy: the careers page, the changelog, the integrations list, a recent blog post, or the pricing tiers. Those pages leak signal because nobody focus-groups them.
When I paste a careers page that lists five open SDR roles, Claude Sonnet 4.6 writes something like 'You're scaling outbound headcount fast, which usually means ramp time is the bottleneck before quota.' That is a sentence a prospect reads and thinks, yes, that is literally my problem. It came from a job listing, not a crystal ball.
- Best inputs: careers page, pricing tiers, changelog, recent blog post, integration list
- Worst inputs: hero headline, mission statement, 'About us' founder story
- Always paste at least a full paragraph, not a single line, so the model has context to reason over
- If the company is tiny and the site is one page, paste the whole page and let the offer-fit logic do the work
How to Tell the Observation Will Land Before You Send
Here is my five-second gut check on every output. Read line one and ask: could this sentence be true of a competitor too? If yes, it is too generic, and I re-run with a different paragraph. The observation has to be specific enough that it would be slightly wrong if pasted to a different company. 'You care about customer success' fails the test. 'Your three newest case studies are all in fintech' passes, because it is verifiably about them.
The prompt has a quiet safety valve I have come to love: if the copy I paste is too thin to support a real observation, it refuses to fake one and asks me for a detail instead. The first time it did this I was annoyed, because I wanted my email now. Then I realized it had just stopped me from sending a generic message to a prospect I actually cared about. A model that knows when to say 'I don't have enough to work with' is worth more than one that confidently makes things up.
The Workflow I Run Every Morning
My routine is dull and that is the point. I open my list of 20 target accounts, and for each one I spend two minutes finding the most revealing page, paste it into the prompt with my offer line, and review the email. I rewrite maybe one in five outputs where the observation misses, and I send the rest. The whole batch takes under an hour, and every email is specific to a human being instead of a mail-merge field.
Batching It So You Don't Burn Out
One more tactical thing that took me too long to figure out: do not write emails one at a time as you find prospects. Context-switching between researching and writing is exhausting and it slows you down. I now batch. First pass, I just collect the best revealing page for 20 accounts into a doc. Second pass, I run all 20 through the Cold Email Written From a Prospect's Actual Website prompt back to back. Third pass, I review and send. Separating the looking from the writing roughly doubled my throughput, because each pass uses a different part of your brain and switching constantly between them is where the fatigue lives.
I'll also say the quiet part out loud: the prompt does not replace judgment. About one in five outputs needs a human tweak because the model latched onto a true-but-irrelevant detail. Your job is to catch that. The model handles the structure and the tone; you handle the taste. That division of labor is exactly right, and it's why the output never reads like a robot wrote it, because a robot didn't, it co-wrote it with a human who'd done the homework.
If you are just starting outbound and drowning in 'just checking in' guilt, start here. The research habit the prompt forces on you is the real skill; the model just speeds up the writing once you've done the looking. Grab the Cold Email Written From a Prospect's Actual Website prompt on Prompt Dock, paste in a careers page instead of a hero headline, and you'll feel the difference in the very first email it writes back.