This one is ours. CODKRAFTERS sells to small businesses in the United States, and the system that finds them and writes to them is ours. Cold outreach is a thing people are right to be wary of, so the honest way to describe this is by what it refuses to do.
Nothing is said about a business that was not measured. Deterministic detectors load the company's own homepage in a phone-sized browser — 390 by 844 — and check it for sideways scrolling, a phone number that cannot be tapped, text too small to read and tap targets too close together. Speed scores come from Google's own PageSpeed Insights, so the reader can re-run them. Every finding carries the words it came from and the page it was found on, and a detector that is unsure returns nothing.
A model writes the email, and that is close to all it does. It is never asked what is wrong with a website — that question produces confident fiction. It is handed the findings and asked for two or three human sentences around them, and a draft citing a defect that no detector found is thrown away rather than repaired. The score that ranks a lead is arithmetic done in ordinary code; the model contributes two of its five components and each one is clamped.
The prompt forbids prices, invented owners' names and claims about lost revenue, and the code does not trust the prompt. Every draft is checked for money, geography, agency language and unsupported commercial claims, and against a word floor and ceiling, because a model cannot measure itself. A draft rejected for length gets one retry. One rejected for inventing something does not.
An address is never constructed. Not from a name, not from a pattern, not from a model: either the business published it or we have nothing — and that is enforced at the point of sending, where no change of configuration can route around it. A site that publishes a notice against having addresses collected yields none. Mailboxes that exist to receive abuse reports are refused outright. Every message carries a postal address and a working opt-out, and a reply using it writes to a suppression list that every later send checks.
Nothing sends on a timer. Drafts are written unapproved and stay that way until a person has read the words, and the approval is checked again at the moment of sending, because days pass in between and somebody may have asked to be left alone since. The cap is fifteen messages a day, beginning at two, with a randomised minute between them. There is no tracking pixel in any email.
It is not a volume machine, and was never meant to be. One mailbox, one sending address, and about forty modules of Python on a small cloud instance that wakes for twenty minutes a night, then shuts itself down. The hard part was never sending email. It was writing the list of things this system may not do into code, where forgetting is not an option, instead of into a policy document.
Built with
This was AI automation for Outreach Engine. See how we approach AI automation, or tell us what you want to build.