A Digital Garden - frequent growth but a bit weedy in the back corner
Just finished an excellent course on Claude Code run by Philipp Acsany over at Real Python. Well, I say I’ve just finished it, in fact the actual course finished late on Sunday evening and it’s now Friday evening. Since then I’ve been going over and fleshing out my notes and trying the material out a bit on a few of my own projects. Given that the course itself was two four hour sessions, you can see it was pretty dense in terms of the information imparted.
The course begins before it begins with some prerequisite setup of things like git, GitHub, UV and Claude Code itself. Philipp was open and responsive to supporting attendees through this process which was a good sign. I had some concerns about a complete lack of experience in agentic coding but it turned out his reassurances were well placed as it really wasn’t a problem.
Speaking of which, there were around forty attendees across most time zones with a broad mix of backgrounds and experience and while I can’t reliably speak for others, it felt to me that everyone would have taken away a substantial amount of learning. One possible improvement might be to have an opt-in buddy up system or some form of chat or Discord outside the Zoom meeting so we could help each other out with the homework or just generally connect. Other than questions and answers in the Zoom chat there was no real opportunity for interaction between the course participants. That would be some nice icing on the cake I think.
The bulk of the material was delivered over Zoom supported by web hosted handouts. It was split into two four hour sessions over a Saturday and Sunday evening here in the UK which worked really well. There’s a lot of solid material to absorb so four hours felt enough without drifting into… ‘drifting’ territory. Daytime on Sunday was a very welcome chance to play catch up with the absorption process and this was supplemented by the optional homework.
In terms of the material, we walked through setting up the bare bones of an application during the first session using this process as a vehicle for exploring Claude Code’s functionality, limitations, gotchas, integration with tools like git and GitHub and the practicalities of its day-to-day usage. This latter point was where I found the most value. You can easily find free material online for learning pretty much any topic but it’s the relating of personal experience that’s worthy of payment. This is a £400 quid course so it’s a reasonably significant personal investment for someone in my position (retired and essentially just having a play for shits and giggles) but it was specifically Philipp’s experience that made it feel like money well spent. As I said in the post-course survey - “No regrets”.
Moving on from the basics we implemented a ‘bug fix’ as a vehicle for layering a generic workflow on top of the bare bones mechanics of Claude Code. I think this workflow, developed into a set of Claude Code skills during the second session, was the biggest tangible takeaway from the course as it represents a really solid base that can be tailored to anyone’s own personal needs. It’s probably a bit overkill for a quick and dirty scripting session and it’s probably missing a few bits of custom specificity for working within an already defined development process. But it seems like a very good jumping off point and as Philipp said, it’s nothing new in terms of it just being a mirror of common development workflows so it’s easy to tweak to taste.
The overnight homework was to implement three small features using the workflow we’d developed during the first session. Iterating this loop three times felt about right in terms of gaining familiarity with the tool and process as well as ironing out the inevitable points of friction and misunderstandings. Personally I spent the whole day working on these tasks and I gained a lot from doing so, but many people didn’t do the homework at all. This wasn’t really an impediment in practice as Philipp ensured we were all on the same page with the same codebase to begin the second session. I do recommend working through the homework if you’ve got the time though as it meant I could focus less on the mechanics of Claude Code and more on the new material presented in the second session.
The second session was a little more disjointed. That’s in no way a criticism. The first session covered the basic usage of Claude Code and felt like a coherent single story from start to finish. The second session felt much more like a collection of short stories. We covered things like
CLAUDE.md file for setting up project-specific or user-specific ‘preferences’ (I use the term loosely cos there’s a good bit more to it than that)Both sessions included a Q&A session at the end and a couple of breaks for a brew, a Pinkelpause, and a smoke.
It’s the first time I’ve really played with AI coding and I have to say I was pleasantly surprised by the process itself as well as the course. My impression was, to a certain degree at least, that AI just kind of produced a shitty first draft. I think I was very much mistaken in that regard. It seemed to reason well and be both perceptive and thorough. Particularly noteworthy was the way it discovered issues that I know for a fact I would not have found without a good deal of effort. And I was also presently surprised by how unopinionated it was. When it found issues or ambiguities my general sense was that it would ask me how I wanted it to resolve these issues rather than just picking an option and doing what it wanted.
Does it produce perfect code? Exactly the way I want it? No. But it does seem very good at implementing changes based on my input as to what I do want it to look like.
Is it early days? Yup. Obviously I’ve only played with this thing and I’m sure I’ll find annoyances. But that’s the same for any tool.
Overall? Happy bunny.
And I got a certificate!