The IRB data warning is one I keep seeing people skip over and it matters. Research workflows have a specific trust problem - you're often working with data that's one mistake away from a compliance issue, and Claude Code's default behavior is to read everything in the project directory.
The context window degradation point is also underappreciated (most people just keep going until outputs get weird). Good that you're building the Markus Academy series around this - researchers need the setup guide more than developers do tbh, the stakes for getting it wrong are higher.
This is such a well-structured breakdown, thank you for putting it together. The ladder analogy hit close to home — I spent months at Level 0-1, convinced I was "using AI" when I was really just copy-pasting between tabs.
The part about context window degradation is something I wish someone had told me earlier. I've had sessions where Claude started giving noticeably worse answers and I thought it was a model issue, not a context problem. That mental model shift changes how you work with it completely.
What you've described here is exactly the gap we're trying to bridge with a course we're building — "Master Claude in the Real World." Not the theory of what Claude can do, but the actual workflows: prompting, Claude Code, Cowork, and integrations. We just launched on Kickstarter for anyone who wants to follow along as we build it out: https://shorturl.at/ZrG8p
Excited to see where this series goes. The jump from Level 2 to Level 3 is where things genuinely start to feel different.
Having worked with cursor recently, I'd say that the setup of terminal etc and workflow including restarting of chats/cleaning up of context windows may be easier to do out of the box in cursor.
Cursor also allows/favors different LLMs, which one can see as a strength or weakness.
I use the Static.app to host my claude code online. I think it's important to know that artifacts could expire, that why a statis website hosting is necessary.
My university offers ChatGPT/codex to all faculty, but none of the Anthropic products. Do you know if they are largely interchangeable for the kind of applications (data analytics, searching for data on the web, spawning, opening, reading, closing files) you describe in this video series?
I don't find cursor that different to claude to be honest, you can ask it to organize your folders and use your terminal as well, the only thing I don't like is that cursor default model doesn't produce very readable code, but I fill claude code is getting a lot of hype when similar tools are almost as good and cheaper.
By typing /compact, Claude will do a summary and determine what to retain; users don't need to do explicit summary before calling /compact, unless advanced users want to implement custom compaction algorithms.
Really appreciate these videos as I try to improve my own Claude Code workflow! One question I had that I haven't seen touched on -- what are your thoughts about using Claude Code in the desktop versus in VS Code versus in the raw terminal, as you seem to do in these videos?
From what I've read online, the Claude Code in the desktop app is disfavored if you're trying to parallelize a lot of agents, but for more basic workflows it seems to work the same as in terminal? I like the way Claude Code in the desktop app feels like a natural extension of using the basic chatbots but I would love to know if there are good reasons I ought to make the switch to pure terminal. Thanks!!
I think you should work however you feel comfortable! I think in the end, doing command line things will be the closest to having a real programmer working on your system, but if you’re more comfortable now in the app, you should start there!
I’m confident that Anthropic will do as much as they can to open up features to desktop app users.
The IRB data warning is one I keep seeing people skip over and it matters. Research workflows have a specific trust problem - you're often working with data that's one mistake away from a compliance issue, and Claude Code's default behavior is to read everything in the project directory.
The context window degradation point is also underappreciated (most people just keep going until outputs get weird). Good that you're building the Markus Academy series around this - researchers need the setup guide more than developers do tbh, the stakes for getting it wrong are higher.
This is such a well-structured breakdown, thank you for putting it together. The ladder analogy hit close to home — I spent months at Level 0-1, convinced I was "using AI" when I was really just copy-pasting between tabs.
The part about context window degradation is something I wish someone had told me earlier. I've had sessions where Claude started giving noticeably worse answers and I thought it was a model issue, not a context problem. That mental model shift changes how you work with it completely.
What you've described here is exactly the gap we're trying to bridge with a course we're building — "Master Claude in the Real World." Not the theory of what Claude can do, but the actual workflows: prompting, Claude Code, Cowork, and integrations. We just launched on Kickstarter for anyone who wants to follow along as we build it out: https://shorturl.at/ZrG8p
Excited to see where this series goes. The jump from Level 2 to Level 3 is where things genuinely start to feel different.
Great series of posts!
Having worked with cursor recently, I'd say that the setup of terminal etc and workflow including restarting of chats/cleaning up of context windows may be easier to do out of the box in cursor.
Cursor also allows/favors different LLMs, which one can see as a strength or weakness.
Great to hear! I should know more about cursor. I actually use pi, which is another coding harness that allows more LLMs
Commenting so I remember to come back when I have time to follow along more closely.
Super useful thanks!
I use the Static.app to host my claude code online. I think it's important to know that artifacts could expire, that why a statis website hosting is necessary.
My university offers ChatGPT/codex to all faculty, but none of the Anthropic products. Do you know if they are largely interchangeable for the kind of applications (data analytics, searching for data on the web, spawning, opening, reading, closing files) you describe in this video series?
I don't find cursor that different to claude to be honest, you can ask it to organize your folders and use your terminal as well, the only thing I don't like is that cursor default model doesn't produce very readable code, but I fill claude code is getting a lot of hype when similar tools are almost as good and cheaper.
Am I understanding the ordering correctly: first ask Claude to write a summary and then run /compact?
By typing /compact, Claude will do a summary and determine what to retain; users don't need to do explicit summary before calling /compact, unless advanced users want to implement custom compaction algorithms.
Really appreciate these videos as I try to improve my own Claude Code workflow! One question I had that I haven't seen touched on -- what are your thoughts about using Claude Code in the desktop versus in VS Code versus in the raw terminal, as you seem to do in these videos?
From what I've read online, the Claude Code in the desktop app is disfavored if you're trying to parallelize a lot of agents, but for more basic workflows it seems to work the same as in terminal? I like the way Claude Code in the desktop app feels like a natural extension of using the basic chatbots but I would love to know if there are good reasons I ought to make the switch to pure terminal. Thanks!!
I think you should work however you feel comfortable! I think in the end, doing command line things will be the closest to having a real programmer working on your system, but if you’re more comfortable now in the app, you should start there!
I’m confident that Anthropic will do as much as they can to open up features to desktop app users.
https://nandigamharikrishna.substack.com/p/the-claude-code-leak-how-512000-lines?r=8op1j&utm_campaign=post&utm_medium=web