Personalized cold emails with sources

A personal email to a named buyer fails in two predictable ways. The research is shallow, so an opening built on a university, a funding round or a job title would work just as well for any of the company's four hundred peers, and the reader sees a template. Or the model fills a gap with something plausible, a figure slightly off or a customer nobody mentioned, and the reader catches it in the first line. A better prompt doesn't prevent either, because the model writing the email is also the one deciding whether it followed the instructions.

Torke Reachout splits that work. Two AI agents do the reading and the writing, where judgement is needed, and code checks everything that can be written as a rule, giving the same answer on every run. If a sentence says 400 and the source it cites says 340, the gate holds it. If a sentence names a company that appears nowhere in its evidence, the gate holds that too.

The node's test files follow a fictional prospect, Lena Brandt, VP Revenue at a fictional company called Example Robotics. Her research holds seven public sources, four written by her. In a post dated 18 September 2026 she wrote that her reps hear about new warehouse projects "from the press, three months too late", and on a June panel she said most lost deals were gone before the team knew there was a deal. The email the gate accepted opens on her own post, names what she wants, which is her team involved before a tender is written, and only then says in one sentence what the sender does. It closes on a single question she can answer in a line, and each sentence carries a tag pointing to its source, so a reviewer checks the whole email in a few seconds.

The code proves that an email is grounded. It cannot tell whether a hook matters to the person or whether the email reads well aloud, so a person on your team reads every draft before anything leaves. What changes for you is that a personal email no longer depends on how carefully someone researched that day: every message carries a trail back to the prospect's own words, and the weak ones stop at the gate.

It never sends and never logs into a website. It won't write to an existing client, an open deal or a competitor, and it refuses any claim about health, family, religion, politics or personal finances, in line with the Swiss data protection rule of collecting only what the purpose needs.

Your company can only describe itself through the lines you approve in one configuration file, and an email can cite a client result only once that result is approved. Tone is measured on the surface, from at least three texts the person wrote, and with fewer it falls back to company copy and says so. A draft that passes every check can still be flat, which is why a person reads it before it goes anywhere.

The research gate, the email gate and the full chain from research to rendered file are written, and the 33 automated tests passed on 15 October 2026, each feeding a known mistake to the checks or confirming that the fictional example passes. No run on a real prospect is recorded yet, so the first prospects go through the full chain with a person reviewing every rendered email.