CRM on Markdown, part 1: how to create a CRM in an evening — step by step, with an AI agent that finds leads

I had three live contacts in three different places: a Telegram thread, a LinkedIn profile, a voice message. And not one place where I could see whose move it was and what I had promised each person. If that sounds familiar — below is a step-by-step guide to getting, in one evening, a CRM that an AI agent runs entirely on its own: it files the contact cards, it tracks who has gone quiet, and it helps you find new leads. No database, no SaaS, no subscription to yet another service — under the hood it's plain text files in a folder.

If your pain is hiring rather than sales: an applicant tracker (ATS) is the same system, only the stage vocabulary changes. Instead of "found → messaged → replied" you get "applied → screening → interview → offer". Everything below applies one to one.

What you end up with

You talk to Claude in plain words — "add a contact", "she replied, here's the text", "what's on fire?" — and it keeps one file per person and builds a board:

The board is forbidden to lie: if the next step is "message them" and there's no exact address, the agent goes no further and the board lights up with NO ADDRESS. If your own move has sat untouched for two days — 🔴. Every one of these rules grew out of a real incident — part two is devoted to them.

Now, step by step.

Step 1. Install Claude — the no-code path or the terminal path

No code — Claude Cowork. This is the Claude desktop app's mode for non-programming tasks: the agent works with files and folders, you just talk. Download the app at claude.com/download and open the Cowork tab. It shipped in January 2026; since July there are web and mobile versions in beta, and since August Cowork also runs in the Chrome side panel — which will come in handy in step 3.

With a terminal — Claude Code. If you're at home in the command line, Claude Code will do — same thing, but in the terminal.

Both options require a paid Claude plan.

Step 2. Add the md-crm skill

A skill is a folder with instructions and templates; once it's in, Claude knows how to run your CRM. I packed it with contact card templates, the board script and the tests.

In Cowork and on claude.ai: download md-crm-skill.zip from the repository, then in the app: Customize → Skills → "+" → upload the archive. The skill library is shared between Claude chat and Cowork — install once, works everywhere.

In Claude Code: two commands inside a session:

/plugin marketplace add d7561985/md-crm
/plugin install md-crm@md-crm

Step 3. Connect the browser — this is about finding leads

A CRM with no inflow of people is a pretty empty table. The inflow here works like this: install the Claude in Chrome extension — it lets Claude see the page you have open. Since August 2026 the Chrome side panel runs full Cowork with all your skills (Max and Team plans get it right away, Pro is being rolled out next).

From there the loop is simple: you open a LinkedIn search and scroll through the results — Claude in the side panel reads what you see, filters out the noise, grabs addresses and files cards into the CRM. That's exactly how mine got populated: one pass — 12 cards.

An important boundary, and it isn't only a legal one: LinkedIn forbids automated scraping, so a human does the scrolling and the agent works with what's already on screen. In practice that's also the right division of labor — decisions about people stay with you.

Step 4. Say: "set up a CRM in this project"

Claude will ask three questions — CRM or ATS, what language for the stages, which folder — and set everything up itself: cards, board, guards. The stage vocabulary is a setting: want it in Russian, in Polish, tailored to your hiring process — just say so.

Step 5. Find your first leads

Ask: "find me leads, here's my niche". The skill knows the procedure; two tricks from it are worth knowing yourself.

Search for posts, not people: the author of a fresh post is active by definition, whereas people search can't tell a live profile from a dead one — my first found contact had 452 followers and zero posts.

And search with pain words, not purchase words: a query built from the vocabulary people use to complain about the problem gave me three live candidates out of three on my first pass, while a "looking for a contractor" query gave one out of three and a lot of self-promotion. One pass and one niche is a data point, not a law — but start with the pain.

For each person found, Claude will file a card with two addresses — the profile for direct messages and the post link for a comment. These are different addresses, and mixing them up is expensive: I once had a prepared message land under a stranger's post.

Step 6. From here on, it's just conversation

"I messaged him" — Claude records what was sent verbatim and moves the ball. "She replied, here's the text" — it records the reply and suggests the next step. "What's on fire?" — it shows the board sorted by urgency. Once a week — "run the board": what's burning, who to ping, who to let go, and one sentence on what got decided this week, into the log.

What the agent ends up taking over

Here is the split after all the steps. The agent files and maintains the cards, remembers every conversation verbatim, builds the board, watches the deadlines, demands an exact address before every send, helps find leads and drafts the messages. Three things stay with you: scrolling LinkedIn, deciding, and hitting "send". Semi-automatic — and that is the value: the machine takes the routine and the memory, judgment stays with the human, and the system physically will not let you lose a contact or a promise.

The system is days old — I built it for myself and I use it myself. The thresholds (2 and 5 days), the stage vocabulary, the hypotheses are settings; you will pick your own. How it was built and which rakes its rules grew out of, with all the numbers and no makeup — in part two.

If you want the same, but without setting it up yourself

The skill is free and open — the md-crm repository. If you want the same thing tuned to your funnel, your hiring process or your team — write to me, this is what I do: I build AI loops for specific businesses.