Notes on working with AI
Research before you reach out
Generic outreach isn't dead because of email. It's dead because it's lazy. One research step gives you a reason only that business would notice.
Same offer, two opening lines
First line one: "I help companies like yours grow with AI."
First line two: "Your site shows three open sales roles and no pricing page, which often means every deal is being quoted by hand. That's the process I help teams shorten."
The second one is an example I made up, but look at what it does. It could not be sent to anyone else. It proves that someone read the page. That difference is the whole idea of research-first outreach, and it has nothing to do with the email tool.
What the research step is
Before a single word of the message gets written, you collect a few facts about the business from public sources: what it sells, who it sells to, what it is hiring for, how it advertises. Then you ask one question of each fact: would this company recognize itself in a sentence about it? The facts that pass become the opening line. The ones that don't are thrown away.
AI is good at the slow part. Reading thirty pages of a company's site, noticing the open roles page and the missing case studies, and drafting a first line that points to one of them takes a model a minute. A person does it for the first prospect and gives up by the fifth.
The developer version: Harvey
Harvey is an open-source sales agent (MIT license, about 3,900 stars on GitHub) that runs on your own machine, using your Claude Pro or Max subscription, or Codex. No API key. It follows the research-first logic step by step:
You train it on your product's website. It proposes the signals worth collecting about a business, and gathers only the ones you confirm. It finds businesses, reads each one's public pages for things like agency use, ad spending or open roles, identifies likely decision-makers and checks that their email address is deliverable. Then it drafts a three-email sequence in which each message refers to something specific about that business.
The drafts wait in an Outbox for your approval. Autopilot exists, but you have to turn it on deliberately, and I would think hard before doing that. Replies are classified and answered with drafts, and anything unclear is escalated to you.
It needs Python and some technical setup, and the project recommends a separate sending domain. It is a tool for a developer or a technical founder, and you do not need it to get the benefit.
Doing it by hand with Claude
The research is a prompt. Turn on web search in Claude and give it one business:
Research this company from its public website: [address]. Find three specific facts about what it does, who it serves, and anything it recently launched or is hiring for. Give each fact with the page it came from. Then draft a two-sentence opening line for a first message from me, about [what I offer], that refers to one of the facts. Mark anything you are unsure about.
It works the same in three places. In B2B sales it finds the reason to write. In recruiting, the same prompt on a candidate's public portfolio or the team page of their current employer gives you a line that says why you picked them. In partnership outreach, it shows what the other company already does, so you can propose something it does not.
The human stays in the loop
What comes back is a draft with sources. You open the sources, delete any claim you cannot confirm, and only then send. A research step that invents a fact is worse than no research.
One more thing. Commercial email in Israel is regulated. Check what you are allowed to send, and to whom, before you send anything at scale.
What to do this week
Pick the one prospect you have been meaning to write to, run the prompt, and read the three facts. If none of them makes you say "that's a reason to write," the business is not yet worth a message.