The AI Sales Stack

The problem

I was the first US sales hire at a global staffing startup. I took over from founder-led sales, which meant inheriting a pipeline that mostly lived in people's heads.

Deals moved in Gmail threads, Slack messages and calls. The CRM only knew what someone remembered to type in. So it went stale fast, and a stale CRM is how balls get dropped.

I was the whole US sales team. I didn't have time to also be my own admin. So I built one.

The system

Four Claude skills, each with one job. They read from everywhere I work. They write to very few places, and every write has a rule.

Reading is safe, so the skills read everything. Writing is where trust breaks, so each skill gets a narrow write path and a gate.

Four workflows

Inbox

Most of my selling happened in email. The Inbox skill reads the whole thread before it says anything. That sounds obvious. But search results only show the first few messages of a thread, and the reply you care about is usually the newest one.

Then it checks Slack and my notes for context, drafts a reply in my voice, and scrubs it for the tells: em dashes and stock phrases. It also answers "where are we with this deal?" with the last touch, who has the ball, and what each side promised.

Touch Logger

The CRM's "last activity" date was the number everyone looked at, and it was always wrong. My real touches happened in Gmail and Slack, where the CRM couldn't see them.

Touch Logger runs a few times a day. It picks up my sent mail, replies, Slack messages and call summaries since the last run. It matches each conversation to a deal or lead by thread, not by single message. Once a thread is tied to a deal, every later reply belongs to that deal, even from someone new. Then it adds one short note per record.

CRM Lifecycle

Every create and update in the CRM goes through one skill. New leads, qualifying a lead into a deal, stage moves, and closing a deal won or lost.

It searches by email and company before it creates anything, so I don't end up with three copies of the same person. A closed-lost deal needs a real reason, because that reason decides who gets a call six months later.

Daily Recap

This is the "never drop a ball" sweep. It reads everything since the last recap: calendar, call recordings, email, Slack, the CRM and e-signature status. Then it groups it all by account.

Each kind of event has a checklist. An external call only counts as done when the CRM has a note, the next step is current, the follow-up email exists, the team heard about it, and every promise I made is a task with a date. Each empty box becomes a numbered ball.

Balls carry over between runs. Anything open three business days moves to the top. When proof shows up later, like a reply or a new note, the ball closes on its own.

Results

I ran US sales alone, and the CRM stayed current without me living in it.

$428K

ARR closed

$1M+

Pipeline built

TODO

Hours saved per week

What I learned

The value was in the gates. Drafting a reply is easy. What made me trust the system was knowing exactly what each skill could change, and that anything else needed my yes.

Read the whole thread. Most AI mistakes I saw came from partial context. A snippet, an old message, a reply it never saw. Forcing full reads fixed more than any prompt tweak did.

Checklists beat memory. A recap that says "here's what happened" is nice. A recap that says "this box is empty, here's the fix" is what kept deals from going quiet.

Judgment stayed with me. Pricing, which deal to push, when to back off a patient prospect, what to say on a hard call. The skills got me to those moments with the context ready. They didn't make the call.

Want this on your team?

I'll build this for your team in my first month. Your inbox, your CRM, your rules for what the AI can touch. Then I'll use it to sell.