i promoted myself to middle manager

last issue i talked about how i'd been saving every email i was actually sending to inbound prospects at openrouter. the inbound, the draft, the version that went out. building a corpus of what good looks like at my company.

i ended that one saying it meant i could teach a robot to do my job.

a few of you wrote back and asked how that actually goes.

so. here's how that goes.

quick context. there's a robot at my company named dove. dove is a slack bot, but really it's a vm running claude code that lives in our slack. i can ask it to do things and i do all the time. but for inbound, i don't have to ask. dove watches the channel where new leads land, and when one comes in, it kicks off automatically.

first it does research on the company. then it drafts a reply: opener, discovery questions, closer. all of that happens before i've even seen the lead.

so the bot's writing emails on my behalf all day. i never see the prompt that produced the draft. i just see the draft.

for a while dove was bad at it. corny openers. "makes sense re:". "two quick ones for you." product features that don't exist. questions that telegraphed their own answers. maybe one in five drafts was something i'd actually consider sending.

so i went and built the corpus.

i had a couple hundred emails saved. for each one i had the inbound message, dove's draft, and what i actually ended up sending. i fed all of it to claude code and said group these into archetypes. what kind of inbound is this. what are the questions i ask. what's the structure i use.

claude found five buckets. net new sales. existing customer. reseller. billing handoff. wrong product. for each one, a tight template: how i open, the discovery questions i ask, and how i close.

on top of that, a small classifier. it decides what kind of email needs to be drafted, and dove drafts that type using only that bucket's template. no general-purpose prompt. no "be a great sdr."

a few rounds of feedback later, almost every draft is something i can send.

i didn't write the rules. the rules were already in my own emails. claude just pulled them out.

here's the part that matters for you. you have a corpus too. every email you've sent. every dm. every cold call transcript. every linkedin message. anything you've done a lot of times in writing, you have receipts for it.

claude is good at turning that into structured data. buckets, templates, the questions you ask, the way you open. you don't have to write a single rule. you just have to point at the work you've already done.

if you want a quick way in, take a folder of stuff you've written. emails, dms, slack messages, whatever. drop it into claude and ask it to write the prompt that would produce work like that.

the result usually surprises you. you've got patterns you didn't know you had. depending on the medium, you can either hand it off to a bot from there or just operationalize it yourself and start sharpening.

leafcutter ants reject any leaf cut more than four percent off the colony's existing samples. oversized pieces get dropped at the entrance and never make it into the fungus garden.

this issue was written with claude code.

if this was useful, forward it to someone who's been hand-editing ai drafts all day. that's the move.

reply here if you want to talk about it. goes to my personal inbox. or hit me up on linkedin if you want a more technical walkthrough.

here's the link to subscribe :)