The GTM Loop

Your first 10 minutes

Before you install anything, see the scoring step work on your own market.

Pick 5 companies you already want to sell to. For each one, find one thing that changed there lately: a job ad, a new leader, a funding round, a tool they use got sold. Then paste this into Claude with your 5 companies under it:

I sell [what you sell] to [who you sell to].
For each company below, I give you one thing that changed there.
Ask three questions of each change:
1. Does it have a date?
2. Can we list who it hit?
3. Does my result fix it?
Answer each question yes or no, with the reason in one line.
Then mark the company KEEP if all three are yes, or WATCH if any one is no.
For every WATCH, say what would have to happen to make it a KEEP.

Companies:
1. [company] - [what changed, and the link where you saw it]
2.
3.
4.
5.

You know this worked when: you have a KEEP or WATCH next to all 5 companies, and you can say why for each one. The KEEPs are who you write to this week. The WATCHes you check again next week.

That is step 2 of the four steps below. The rest of this page is how to make it run every week.

What this is

Most AI sales setups are a straight line. Find people, write, send, done.

This is the loop version. Nine steps, an agent on each one, and a tenth step that takes what worked and feeds it back into step 1. Next month's research starts from this month's sales calls.

Two things are in here that the other builds in this vault leave out:

1. The whole install. Every account, every key, which connections to turn on inside Claude, and where a connection stops short so you have to call the tool's API directly.

2. What breaks, and the guard for each one. Most of these failures are quiet. Nothing errors. The numbers just come out wrong, or nothing sends, and it looks fine.

The four steps from the post

The post names four steps. Here is each one, the way we run it. The full build further down adds the install, the keys and the guards.

1. Sort the market into three groups. Every week, look for changes in three places.

• What the market does to the buyer. A tool they rely on gets sold.
• What the buyer is doing. A job ad for the problem you fix, a new leader, a funding round.
• How their market is changing. A trend with no date and no names.

Never write to anyone over a trend alone. The trend shapes what you say. The first two groups pick who you write to.

2. Score every change before you write. Ask three questions of each change, and write the answers down so two people get the same score.

• Does it have a date?
• Can you list who it hit?
• Does your result fix it?

A change that scores low goes on a watch list, and you check it again next week. Nothing gets thrown away. Move fast on a new leader. Most new leaders pick their tools and vendors in their first 60 to 90 days.

3. What happens to every reply. Each reply gets a tag before anyone answers it: interested, wants info, wrong person, or out of office. Each tag has one written answer, and the agent sends the answer that matches the tag. When the reply is an objection, sort it by what is behind it: trust, money, or another vendor. Keep one written answer per objection. The first time a new objection shows up, write its answer and add it, so the next reply finds it.

4. What feeds back. Every Monday, paste the week's objections and call scores into 11-weekly-loop-prompt.md. It pulls out the exact words buyers pushed back with and writes them into your business file. Step 1 starts from that file, so next week's first messages answer those doubts before anyone asks.

Where each one lives in the build: steps 1 and 2 are Steps 1 to 6, step 3 is Steps 7 and 8, step 4 is Step 10.

When you build the full loop

Everything from here down is the full build: all ten steps, the agents, every account, the files, and what breaks. You do not need it for your first 10 minutes. The links to other builds in this vault are next steps, not required.

The loop

The order is the product. Step 10 is the one almost everyone skips. Without it you run the same loop with the same mistakes every month and call it consistency.

The 13 agents and the brain they share

Every agent below is a Claude skill: one folder with its instructions, its rules and its examples. None of them start from zero. Each one reads the brain first.

The brain is one folder of plain notes (Obsidian works well) that every agent can search: what buyers said on calls, which reasons for writing got replies, which offers booked calls, and what went wrong last time. Ours holds 1,237 notes with 3,955 links between them. Yours starts with one file about your business and grows every week.

The loop closes through the Metrics Agent. It measures every other agent: which reason for writing got replies, which offer booked calls, how each call scored. It writes those numbers back into the brain every week. Next week, every agent above reads them before it moves.

What you need

Install in this order. Each one unlocks the next step.

The full detail for each account: what key to make, what permission it needs, which connection to switch on, and how you know it worked is in stack-install.md.

Get the files

Everything is in a public folder on GitHub. No account, no sign-up.

The folder: https://github.com/gary-chakraborty/gap-build-vault/tree/main/builds/gtm-loop

Build it

Before step 1: one file about your business

Use the business block from The Eight Agents (claude-md-snippet.md) or The Context File. Fill every bracket. Every agent in the loop reads it first, and step 10 is the step that keeps it current.

You know this worked when: a new chat answers "what do I sell and who buys it" with your real offer and your real buyer.

Steps 1 to 6: research, offer, list, send

Run them from The Eight Agents and Signal Check, in order. Two additions from this build:

1. Before you lock the offer, paste 03b-competitor-ads-prompt.md. It sorts competitor ads by how long they have been running. An ad live for 60 days or more, with 3 or more versions, is one that is paying for itself.
2. Before you load anything into a sending tool, read the Smartlead and HeyReach sections of failsafes.md. Both tools can accept your list, report success, and send nothing.

You know this worked when: you can name, for every person on the send list, what changed at their company and when, and your sending tool shows real sends after the first hour.

Steps 7 and 8: reply and book

Set up The Reply Desk and Reply to Booked Call. Then wire the reply alert using the pattern in connections.md: reply → a small relay → Slack.

You know this worked when: you reply to your own test email and a Slack card shows up within a few minutes, with the draft answer already on it.

Step 9: the call

1. The morning of the call, paste 09-call-brief-prompt.md with the person's name, company, the message they replied to, and their reply.
2. Record the call (with permission). Transcribe it with Whisper.
3. Paste the transcript into 10-call-score-prompt.md.

You know this worked when: the score names one exact moment and gives you the line you should have said, not a general tip.

Step 10: feed it back

1. Every Monday, count last week. Use the scoreboard script from Reply to Booked Call (weekly_scoreboard.py) or count by hand from the tools.
2. Paste the numbers into 11-weekly-loop-prompt.md.
3. Make the changes it gives you. Anything that clearly won goes everywhere that same week. Only a thing you are not sure about gets changed on its own, so you can tell what moved. Update your business file so step 1 starts from it.

You know this worked when: next week's list, offer or opening line is different because of something last week's numbers showed.

What breaks first

The five that come up most, from running this loop for real. Full list in failsafes.md.

What is not in here

Your scripts. The templates hold the shape. Your words come from step 2.

A done-for-you setup. This is the full map. Putting it in for you is a different thing.

Tell me how it went

No email needed, nothing to unsubscribe from.

Two minutes, three things: how you found this, whether it was useful, and what you want me to build next. https://form.jotform.com/262194392563059