Overview

Claude Code Install + Setup Call - Friday 24 April 2026, the week before the live Claude Code Training on 30 April. The call covered the setup layer for Claude Code from scratch: installing on Mac and Windows, the startup and wrap-up session discipline, how GitHub fits into a marketer's workflow , running across multiple computers, terminal choice and the multi-session workflow , the two memory layers (auto and manual), PipeBoard for Google Ads, Meta, and TikTok, write-capable MCPs and how not to blow up an ad account, transferring context out of ChatGPT, the cost reality of running Claude at scale, and the mindset shift that makes Claude Code a one-way door.

Summary of the call

Pre-training vs live training. The pre-training modules are the setup layer - install, project memory, permissions, the memory layer, the mental model, session discipline, GitHub publishing. The live training on 30 April is the execution layer - reports, campaigns, creatives, audits, landing pages, emails. Module 04 is the biggest unlock before the live training. Modules 06-07 is the ideal stretch goal. If you walk in without setup done, you'll watch instead of build.

Install paths differ by OS. Webprofits is fully a Mac company - all our internal install material was developed on Mac. Windows works but it's a different path: PowerShell as the terminal, Node.js as a dependency. One member hit a "Windows is not supported by this script" error from the default install command - the fix is going through the Install Claude Code section in Skool which has the Windows-specific path, and the Node.js install. Fastest unblock for any Windows error: post the screenshot in the community thread, the Webprofits dev team will get back to you the same evening.

"Claude is a bunch of MD instructions." Once you understand that everything Claude Code does is driven by markdown files - your CLAUDE.md, your skills, your memory - the rest of the system makes sense. The session has no long-term memory of its own to rely on. Everything that matters has to be loaded in or saved out, deliberately.

Every Claude Code session has the same anatomy. Run the startup skill at the start (pulls memory and context from Supabase + past sessions). Do the work. Run the wrap-up skill at the end (saves everything back). The startup skill imports everything from the previous wrap-up, so the loop only works if you run both, every time. Skip the wrap-up and you lose all that context. Everyone makes this mistake at least once

• you'll close the browser tab, kill the session by accident, and realise you didn't save. Flash Resume can sometimes recover a session, but the safer rule is: never end a session without wrap-up.

Two memory layers - auto and manual. Claude does have an automatic memory layer that saves things agentically as you work, but it's unreliable and won't give you a strong handover into the next session. The manual layer is what the wrap-up skill saves - structured, deliberate, complete. This is one of the few times you give Claude an explicit command rather than a goal. Most prompts describe the outcome you want; "wrap up" is a directive that says exactly what to save and how. The default 50-line read on local MD files is also why we move memory into Supabase - by week two, the core MD files were running into hundreds of lines and context was being missed.

Working in a folder, not on your computer. Always launch Claude inside the project folder you intend to work in. The workflow: copy the path from Finder, type claude in your terminal, paste the path. You're now in that folder. Then run /startup, do the work, run wrap-up. There are multiple ways to get into a folder - what matters is being in the right folder before you start. GitHub is a publishing layer, not a developer tool. Marketers connect GitHub to back up work and share live links to reports, briefs, dashboards, and campaign assets without needing a hosting account. A personal GitHub account is enough - no enterprise tier, no organisation, no paid plan needed. Reports, dashboards, and creative briefs all need to be shared with someone eventually (designer, client, account manager). GitHub gets you a shareable link with no extra plumbing. Webprofits also runs shared client repos and a WP skills repo - the directors own the skills, the team builds against them. Working in the same repo at the same time has been less of a problem than feared. If a repo gets destroyed, you spend half an hour rebuilding it. Don't get attached to anything you can rebuild.

Two-computer workflow (home + work). The hack path is Dropbox sync for files plus a shared Supabase memory database that both computers point to. The proper path is GitHub for files plus the same Supabase database. Alex's actual setup: all three. Dropbox AND GitHub AND Supabase. Why? Because GitHub doesn't capture .env files and other .gitignored files, and you don't always know what's been excluded until you need it. Dropbox syncs every minute as a safety net. Tell your team: build the working folder inside Dropbox or Google Drive (either works), so losing the computer never costs you the work. Alex now also backs up to an external hard drive - context is the superpower, and the compounding effect over weeks is what would actually hurt to lose.

Terminal choice and the multi-session workflow . The reason Warp (Mac and Windows) and Superset (Mac only) matter isn't aesthetics - it's the tab structure. Once you start using Claude Code, you'll quickly hit the bottleneck of waiting for Claude to finish. The fix is running multiple sessions in parallel. Default terminal apps don't handle that. Warp and Superset both do. Alex pushed it to 10 sessions at once - smashed the RAM, computer froze on a video call. Manage the count, but the workflow itself - parallel sessions across multiple tabs - is what unlocks the speed. Developers already in Cursor or VS Code can stay there. Claude Code works in all of them. The course teaches the marketer-first path because that's the fastest route for non-developers. Superset doesn't run on Windows, so Windows users go with Warp. Restart your computer daily. Claude Code generates large temporary files during creative analysis, video processing, and image handling. They build up fast. Restart clears them. Webprofits generated 1.1 million lines of code last month - the load on a single laptop is in a different category from the old browser-tab workflow . If things feel slow, restart before debugging anything else.

The MCP hierarchy. Three categories. Claude.ai connectors are the easiest path - they transfer across environments. Direct platform MCPs (the vendor provides one, you grab the API key, you connect). Third-party community MCPs - be skeptical and know what you're installing. An MCP is essentially an API designed so the LLM knows exactly how to use it - what data is available, how to read and write.

PipeBoard for paid media. Currently the best MCP for publishing campaigns to Google Ads, Meta, and TikTok. Setup flow: sign up at PipeBoard, go to your Claude organisation settings, add the custom MCP via the connectors panel, then go into personal settings → connectors → click connect. In your terminal inside Claude Code, run /mcp to see all connected MCPs. Anything that says "needs authentication" - hit enter, hit authenticate, log in via the browser, restart the session. Now any prompt can audit, query, or write to your ad accounts directly. Critical safety: PipeBoard can read AND write. With auto-mode on plus a vague prompt, it can change budgets, pause campaigns, edit copy. Concrete failure mode: if you have access to six accounts and you don't specify which one, it might update budgets on the wrong account. Always start an account session with a read-only directive: "audit the account and show me areas I could improve, do not change anything". Iterate from there. When you do want to write, be explicit about what to change and where.

"If you have PipeBoard, do you still need a media buyer?" One member asked this directly. Answer: depends if you know how to media buy. Someone has to tell Claude what to do, interpret the data, and make the next-step decision. That can be you. The thing PipeBoard collapses is the mechanical work - launching a Meta campaign used to be horrible (primary text, images, settings, every detail manually) and is now a prompt. The bottleneck on Meta is creative, not buying - audience is broad anyway. So you still need someone making creative decisions and judgment calls. Ben's add: anything that's not very good already should be replaced one way or another. With paid search specifically , if you know what you're doing, you can transform the way media is delivered.

Transferring context out of ChatGPT. Project by project. Tell ChatGPT to consolidate the learnings, memory, and important context from one project into a knowledge file. Export it. Drop it into the equivalent Claude project folder as a reference. Then do one final pass: ask ChatGPT for everything it knows about you, how you like to work, your preferences. Export that too. Then turn ChatGPT off and don't go back. Inside a few weeks of working in Claude Code, you'll be so far past whatever ChatGPT had on you that the migration won't matter.

Rebuild, don't automate. Every existing account, every existing project - treat it as a fresh start. Like starting a new business or a new job. The biggest mistake is putting Claude Code on top of your old analog workflow - all that does is make one step faster while you stay stressed. Rebuild from the foundations: research, context, channel data, competitors, business specifics. The setup is slower at first, then it compounds. The people getting the biggest lift from Claude Code aren't automating - they're rebuilding.

Excellence is the only position that still pays. Average media buyers, average creatives, and average operators are getting replaced by marketers running Claude Code. If you're paying for average work, stop. Look at who around you isn't in Claude Code yet - they will move very fast right now because the value is so visible. Claude Code does not replace strategy, creative direction, or judgment. It multiplies whoever is using it. Bad decisions become bad decisions at scale. The number has to go up. If the number's not going up, who cares about Claude Code. With Claude Code, the number goes up.

Token optimisation matters once you scale up. Webprofits last month: $4,500 in subscription costs (everyone on the top account) plus another $4,500 in additional tokens. $9,000 total for Claude. Better investment than anything else, but there's enormous waste in how people use it in the first few weeks. The Token Optimisation module in the classroom is where that gets addressed - file structure, context loading, all the patterns that compound at scale.

Screenshots are the fastest debugging tool. When Claude hits an error you don't understand, screenshot the error, paste the image back into Claude Code, ask it to diagnose and fix. Works for install issues, authentication failures, permission prompts, terminal messages. The Windows install error in this call was solved this way.

Watch out for the Dunning-Kruger curve. Day one of Claude Code: everyone feels confident. Two weeks in: the valley of despair when you realise how much you don't know. Don't fir e your external media buyer or SEO agency in week one. Don't take everything in-house just because you can. There's a real difference between "I can do this" and "I know which parts I can't do yet". Take it in-house where you can; keep specialist support where you can't yet.

Claude Code is a one-way door. Nobody who gets properly set up goes back. Once you're in, everything outside of it - the old tabs, the old tools, the old workflows - feels fiv e times slower. The fastest thing you used to do is now the bottleneck. Webprofits has had hard conversations with team members about adopting it: it's like running a Ferrari against jet engines. Three months from now, anyone not in Claude Code will be in trouble - the conversations are everywhere, the gap is widening fast.

Why we're confident in this training. Even as we figur e all of this out, even still as marketers (not developers), we can do dramatically more than before. The work is amazing, the client results are great. Ben's view: this training is worth 10x what we've priced it at. The Academy is here to keep sharing whatever we learn next as the platform evolves.

Action items

☐ Get Claude Code installed on terminal (Mac or Windows). Post a screenshot in the community thread if you hit a Windows install error - dev team will respond same evening ☐ Install the startup and wrap-up skills from the Quick Start Guide in Module 00, then restart your session

☐ Work through Modules 01 to 04 before the live training on 30 April (06-07 if you can stretch)

☐ Set up a single working folder in Dropbox or Google Drive (or GitHub) that syncs to both your home and work computers

☐ Sign up for PipeBoard, add the Google Ads, Meta, and TikTok connectors via Claude.ai, then run /mcp + authenticate inside Claude Code

☐ Restart your computer daily - especially if Claude Code starts feeling slow

☐ If you're transferring out of ChatGPT, do it project by project and consolidate each into a knowledge file

☐ Always start a paid-media MCP session with "do not change anything" until you're ready to write

☐ Post any setup questions in the community thread - someone else is hitting the same wall

Next: Live Claude Code Training - Thursday 30 April 2026, 10am - 2pm AEST. We may run another Install + Setup Call next week depending on how the support questions go.

Exercises

1. Review the concepts covered in this lesson: Overview.
2. Write down your key takeaway from this lesson.
3. Practice running any commands or prompts mentioned above inside your terminal.