Why I Built My Own Sales Coach

We had no problem booking intro calls. Converting them was a different story. So I wired Claude into my meeting transcripts and let it tell me exactly why.

The Funnel Was Working

My co-founder Isaac and I are building Dealwire: custom AI tooling for real estate firms. Our plan: find painful, valuable problems and deliver bespoke solutions, tailored to our clients' workflows. First step, let's get some leads.

Within weeks, the top of our funnel was humming.

Every morning, I'd run the same playbook: Hammer LinkedIn outreach. Connect with 30 likely decision makers at real estate firms. When they accept, my Garmin watch buzzes and I DM them while they're still online: short, specific, relevant to their firm. On the cold email side, same energy.

We were getting a 5.1% and 22% response rate on email and LI, respectively. Strong start. People wanted to hear the pitch. And we had no problem booking intro calls: 30 to 40 a month with firms managing plenty of real estate.

Prospects were engaging with us, asking good questions, even offering intros to their networks.

We were finding painful and valuable problems, and building relationships with some incredible people.

But we weren't converting as well as we could. I knew something was off in how I was running calls, but when you're doing several calls a day, each one blurs into the next. You hang up and it's eyes up, next target.

I needed a feedback loop. So I built a coach.

The Setup

I use Granola for all my meetings. It records, transcribes, and gives me structured notes after every call.

The idea was simple: after every sales call, pipe the transcript to Claude and have it grade my performance.

A single rubric wasn't going to cut it. An intro call where I'm trying to dial in on the problem has completely different failure points than a follow-up where reviewing a proposal. So the skill auto-detects the call type and runs a different analysis framework for each. Intro calls get graded on positioning, competitive awareness, and qualification. Process calls get graded on demo effectiveness, objection handling, champion building, and deal progression.

I built it as a Claude Code skill, a single markdown file that connects to Granola via MCP to pull transcripts automatically. I'd type /sales-coach after a call and get a letter grade, call type, direct quotes of where I messed up, and action items. All within a minute of hanging up.

Then I automated it. A cron job runs every 30 minutes during the workday, checks Granola for new meetings, and runs the analysis if it finds one. Then, I get a macOS notification with a txt file when a new call is graded.

What It Found

The first time I ran this across a batch of 10 calls, it was brutal. But it was also the most actionable feedback I'd gotten: specific, grounded in my actual statements, and delivered within a minute of hanging up. That immediacy is the whole point. When you get feedback while the call is still fresh in your head, you can actually hear what the AI is talking about.

Here's what my AI coach surfaced:

I was asking good questions but pulling up short on our value.

I was actually pretty good at getting people to open up. I'd ask about their workflows, their pain points, what they'd tried. But the AI caught something I couldn't see from inside the conversation: I was spending so much time asking questions that I never got around to explaining what we actually do. I'd have these 20- to 30-minute calls where the prospect talked for 20 minutes, I learned a ton about their business, and then we'd run out of time before I could walk them through how we'd actually help. The call would end with something like "this is really interesting, send me something" and no follow-up booked.

I wasn't helping people sell us internally.

On follow-up calls, I kept hitting the same wall: my contact would say they like what we're doing, but they need to run it by their partner or bring it to a team meeting. And I'd say something like "great, I'd love to connect with them next week." The AI started flagging these moments. It pointed out that I was expecting my contact to go pitch their boss with nothing. My response to "I need to show this to my team" was always some version of "happy to help however I can," which the AI pointed out isn't actually help.

What Actually Changed

I stopped letting good conversations run too long. The AI kept flagging calls where I asked great questions for 25 minutes and never explained what we do. So I started giving myself a mental halfway mark. By this point, I need to transition from "tell me about your process" to "here's how we'd help." I'm still not great at it, but I'm aware of it now, and that's the difference.

I focused on making it easy for people to sell us internally. Whenever someone says "I need to bring this to my partner," I used to just say "sounds good." Now I try to send them something concrete: a short summary with their specific numbers and what the first few weeks would look like.

Why This Works

Getting feedback within a minute of hanging up, while the call is still playing back in your head is completely different than watching sales training videos on Sunday afternoon. You remember the exact moment. You remember what you were thinking when you said the wrong thing. That's when the feedback began to rewire my internal framework.

Before, my feedback loop was basically non-existent. I ran a call, and moved on. By the time I thought about what I could've done differently, the details were gone and I was already on the next one.

Now I'm learning in something close to real-time. I'm not going to pretend I've figured out our process; I'm still making mistakes on every call. But I'm making different mistakes, and I'm catching patterns faster than I would on my own. The AI doesn't get tired of telling me the same thing. If I keep falling into the same habits, it keeps flagging them until something clicks.

The other thing: it's honest in a way that people usually aren't. Isaac is an absolute mensch, but he isn't sitting on every call taking notes on what I should've said differently. The AI just tells me.

My Takeaway

I'm not going to pretend this fixed everything overnight. But the gap between "great conversations" and contracts narrowed quickly when I had something telling me with specific quotes exactly where I was losing people.

If you're a founder doing your own sales and you're not analyzing your calls, you might just be flying blind. But this is an easy way to get honest feedback on your efforts.

If you want to build your own version, here's the full Claude Code skill and the cron script that automates it. Drop the skill in your ~/.claude/skills/ directory, connect Granola's MCP server, and add the cron to your crontab.