Don’t overwork your reps,
let machines handle it

S/001

From 9 deals to a full ABM program in 3 months

ABM program3 months

A client had no list of who to sell to. The audit told the story: a CRM with no scoring and no tiering, an ICP field empty on every record, and 0.19% of contacts in a sequence.

I rebuilt the targeting from their own wins. I studied their nine closed deals, worked out what a good customer looks like, then found every company like that in the market. Leads scanned at events now flow into the same system, no hands needed.

Routing drawn on the cover, fig. 001.

The same build now tracks buying signals on every account and routes them to the right rep as tasks and Slack alerts.

Three months in, every account carries a tier and eight live signals, and the rep who owns it hears about buying moments the day they happen.

S/002

Full GTM rebuild: TAM, signals, sequences

Clinical AI18 reps3 offices

A clinical-AI company had 18 reps in three offices selling on a platform that fought them: no parallel dialer, emails dying in a broken integration, and a target list nobody trusted. I had to rebuild the whole engine in parallel while the team kept selling on the original.

The history came out first: years of call recordings were trapped inside with no export button, so I found a hidden way in and pulled about 9,000 files with one command. Along the way I caught a cleanup step that was about to rename real drug companies to Spotify. The rebuild took three months and runs nine systems deep:

Now a new trial registration becomes a Slack ping, a sequence enrollment and a rep task before anyone has read the news. Clay, Apollo and Zapier carry the admin; the reps just sell.

S/006

Context engineering: a self-maintaining system for AI agents

Not client work, my own system

Sessions are disposable, files are the memory. I got tired of re-explaining the same client to an AI agent every morning, so I wrote the context down instead. Every client gets two files: one durable with the IDs, the architecture and the known traps, one volatile with what’s in flight, date-stamped. A session starts by reading both, and I explain nothing.

Anyone can gather open tasks. The hard part is knowing which ones are already done: two older routines skipped that check and kept handing me back work I had already finished, plus a client that wasn’t mine. Both are off. The one I replaced them with won’t put a task in front of me until it has checked the task is still open, and it still stages every action for my okay until I have watched it long enough.

There is no framework under any of this and no database, just plain markdown and two shell scripts. Writing to those files is destructive on purpose. A fact that gets contradicted is deleted, not appended under the old one, closed loops get cut, and every file has a line ceiling that pushes the overflow into a topic file beside it. One fact lives in exactly one place. Above all of it sit rules the agents can’t argue their way out of: nothing sends, posts, writes to a CRM or pushes to git without me saying so, campaigns never go live on their own, and credentials can’t be staged into a commit.

What I get out of it is plain: I can hand an agent a client it has never seen and the work comes back the way I would have done it myself. Three of these skills I adapted into my team’s shared repo, along with the headless runner that executes them on a schedule, and the team uses them now.