Master Claude Code

Hey,

Last Thursday was W7: Master Claude Code. The capstone. Seven workshops in, and the series did something different. It stopped talking about AI in a browser tab and started running it inside the repo, live, in front of the room.

The browser-tab ceiling

Here is the pattern you have been living in. Open a tab. Type a prompt. Copy the answer. Paste it where the work happens. Fix what did not fit. Repeat, dozens of times a day.

For six workshops we built real systems inside that loop. A judgment system that reviews work the way you do. A workflow chain that runs one input through multiple roles. A decision framework that stress-tests an idea before you commit. All of it lived in a tab you opened by hand, every single time.

That tab is the ceiling. W7 broke it.

What we shipped, live

No scripted scenario. No problem I made up. We sourced a real pain, live, from Reddit, and shipped a real app for it, in the terminal, while you watched.

1. Source. A scanner pulled real posts from subreddits where people complain about real work. It came back with 198 posts across nine subreddits in about two minutes. Sixty-six qualified as high-signal pain points.
2. Rank. Claude Code read the raw output and ranked the top opportunities itself, with the quote that proved each one. The winning pick: an Excel data cleaner for non-technical users drowning in messy spreadsheets.
3. Ship. It spec'd the build, wrote a single-file web app with no backend and no login, tested its own work, and ran a deliberately messy CSV through it. Three duplicate rows removed. Four values standardised. Three blank rows dropped. A real app, for a real pain, built live in minutes.

Then it built the landing page to collect emails for it. An MVP proves the build. An email capture proves the demand.

Why this only works inside the repo

You could try the same loop in a browser tab. You would copy the scan output, paste it into a chat, copy the ranking, paste it somewhere, ask for the spec, copy that, paste it into a code window, and stitch the pieces together by hand. Every step leaks context. Every round trip is on you.

Inside Claude Code the loop collapses. The output is a file. The ranking reads that file. The spec sits next to it. The build runs against the spec. Nothing is pasted, because the context never leaves the project. That is the entire point.

CLAUDE.md: your judgment, permanent

The W6 judgment system worked, but you had to paste it every time. CLAUDE.md is the file that fixes that. It lives in your repo. Claude Code reads it automatically, before it does anything. Your standards, your conventions, your commands go in once. Then every change, every review, every build carries your judgment without you.

The moment that landed it: on stage, Claude Code built the app in the wrong folder, because I had never told it where to work. One line in CLAUDE.md (always work inside the project directory unless told otherwise) and the problem never repeated. Every surprise in the output is a missing line in CLAUDE.md.

Take it home

Two things to run after W7.

First, open Claude Code in a project you care about, write a CLAUDE.md with the four blocks (what the project is, your standards, your conventions, your commands), and ask it to build one small thing. Watch it plan, test, and ship. That is how tonight becomes real.

Second, the W7 write-ups land on the blog this week: the full source-rank-ship loop, the browser-tab ceiling, and the CLAUDE.md starter template.

Read the W7 write-ups

And the thing I am most excited about. REPOSITION Labs is opening. A community. Weekly sessions deeper than tonight. A library of the systems and apps we build, including the scanner and the Sheet Cleaner from this workshop. Free for now. If you want in, the waitlist is below.

Join the Labs Waitlist

Reply if you want in. I read every reply.

Talk soon,

Merryl