6 Weeks of Claude Code

> 1. Monorepos can save time

Yes they can save you some time, but at the cost of Claude's time and lots of tokens making tool calls attempting to find what it needs to find. Aider is much nicer, from the standpoint that you can add the files you need it to know about, and send it off to do its thing.

I still don't understand why Claude is more popular than Aider, which is by nearly every measure a better tool, and can use whatever LLM is more appropriate for the task at hand.

> 2. Start with a good spec. Spend enough time on the spec. You can get AI to write most of the spec for you, if you provide a good outline.

Curious how you outline the spec, concretely. A sister markdown document? How detailed is it? etc.

> 3. Make sure you have tests from the beginning. This is the most important part. Tests (along with good specs) are how an AI agent can recurse into a good solution. TDD is back.

Ironically i've been struggling with this. For best results i've found claude to do best with a test hook, but then claude loses the ability to write tests before code works to validate bugs/assumptions, it just starts auto fixing things and can get a bit wonky.

It helps immensely to ensure it doesn't forget anything or abandon anything, but it's equally harmful at certain design/prototype stages. I've taken to having a flag where i can enable/disable the test behavior lol.

Agreed, for CC to work well, it needs quite a bit of structure

I’ve been working on a Django project with good tests, types and documentation. CC mostly does great, even if it needs guidance from time to time

Recently also started a side project to try to run CC offline with local models. Got a decent first version running with the help of ChatGPT, then decided to switch to CC. CC has been constantly trying to avoid solving the most important issues, sidestepping errors and for almost everything just creating a new file/script with a different approach (instead of fixing or refactoring the current code)

In my experience, if you don't review the generated code, and thus become proficient in C# enough to do that, the codebase will become trash very quickly.

Errors compound with LLM coding, and, unless you correct them, you end up with a codebase too brittle to actually be worth anything.

Friends of mine apparently don't have that problem, and they say they have the LLM write enough tests that they catch the brittleness early on, but I haven't tried that approach. Unfortunately, my code tends to not be very algorithmic, so it's hard to test.

I'm on the tail end of my 35+ year developer career, but one thing I always do with any LLM stuff is this: I'll ask it to solve something generally I know I COULD solve, I just don't feel like it.

Example: Yesterday I was working with an Open API 3.0 schema. I know I could "fix" the schema to conform to a sample input, I just didn't feel like it because it's dull, I've done it before, and I'd learn nothing. So I asked Claude to do it, and it was fine. Then the "Example" section no longer matched the schema, so Claude wrote me a fitting example.

But the key here is I would have learned nothing by doing this.

There are, however, times where I WOULD have learned something. So whenever I find the LLM has shown me something new, I put that knowledge in my "knowledge bank". I use the Anki SRS flashcard app for that, but there are other ways, like adding to your "TIL blog" (which I also do), or taking that new thing and writing it out from scratch, without looking at the solution, a few times and compiling/running it. Then trying to come up with ways this knowledge can be used in different ways; changing the requirements and writing that.

Basically getting my brain to interact with this new thing in at least 2 ways so it can synthesize with other things in your brain. This is important.

Learning a new (spoken) language uses this a lot. Learn a new word? Put it in 3 different sentences. Learn a new phrase? Create at least 2-3 new phrases based on that.

I'm hoping this will keep my grey matter exercised enough to keep going.

After 16 years of coding professionally, I can say Claude Code has made me considerably better at the things that I had to bang my head against the wall to learn. For things I need to learn that are novel to me, for productivity sake, it’s been “easy come; easy go” like any other learning experience.

My two cents are:

If your goal is learning fully, I would prioritize the slow & patient route (no matter how fast “things” are moving.)

If your goal is to learn quickly, Claude Code and other AI tooling can be helpful in that regard. I have found using “ask” modes more than “agent” modes (where available) can go a long way with that. I like to generate analogies, scenarios, and mnemonic devices to help grasp new concepts.

If you’re just interested in getting stuff done, get good at writing specs and letting the agents run with it, ensuring to add many tests along the way, of course.

I perceive there’s at least some value in all approaches, as long as we are building stuff.

Before AI, there was copy paste. People who copied code from Stackoverflow without understanding it learned nothing, and I saw it up close many times. I don't see a problem with you asking for advice or concepts. But if you have it write everything for you, you definitely won't learn

That being said, you have to protect your time as a developer. There are a million things to learn, and if making games is your goal as a junior, porting GDscript code doesn't sound like an amazing use of your time. Even though you will definitely learn from it.

A few years ago there was a blog post trend going around about “write you’re own x” instead of using a library or something. You learn a lot about how software by writing your own version of a thing. Want to learn how client side routing works? Write a client side router. I think LLMs have basically made it so anything can be “library” code. So really it comes down to what you want to get out of the project. Do you want to get better at C#? Then you should probably do the port yourself. If you just want to have the ported code and focus on some other aspect, then have Claude do it for you.

Really if your goal is to learn something, then no matter what you do there has to be some kind of struggle. I’ve noticed whenever something feels easy, I’m usually not really learning much.

I think this is a really interesting point. I have a few thoughts as a read it (as a bit of a grey-beard).

Things are moving fast at the moment, but I think it feels even faster because of how slowly things have been moving for the last decade. I was getting into web development in the mid-to-late-90s, and I think the landscape felt similar then. Plugged-in people kinda knew the web was going to be huge, but on some level we also know that things were going to change fast. Whatever we learnt would soon fall by the wayside and become compost for the next new thing we had to learn.

It certainly feels to me like things have really been much more stable for the last 10-15 years (YMMV).

So I guess what I'm saying is: yeah, this is actually kinda getting back to normal. At least that is how I see it, if I'm in an excitable optimistic mood.

I'd say pick something and do it. It may become brain-compost, but I think a good deep layer of compost is what will turn you into a senior developer. Hopefully that metaphor isn't too stretched!

Well I think you've identified a task that should be yours. If the writing of the code itself is going to help you, then don't let AI take that help from you because of a vague need for "productivity". We all need to take time to make ourselves better at our craft, and at some point AI can't do that for you.

But I do think it could help, for example by showing you a better pattern or language or library feature after you get stuck or finish a first draft. That's not cheating that's asking a friend.

Have you tried context7 MCP? For things that are not mainstream (like Javascript, Typescript popularity), LLM might struggle. I usually have better result with using something like context7 where it can pull up more relevant, up to date examples.

In my experience LLMs are notoriously bad at tests, so this is, to me, one of the worst use cases possible.

Was it Sonnet or Opus? I've found that Sonnet will just change a few small things, Opus will go and do big bang changes.

YMMV, though, maybe it's the way I was prompting it. Try using Plan Mode and having it only make small changes.

Is this the one that goes 'Oh no, I accidentally your whole codebase, I suck, I accept my punishment and you should never trust me again' or is that a different one?

I seem to remember the 'oh no I suck' one comes out of Microsoft's programmer world? It seems like that must be a tough environment for coders if such feelings run so close to the surface that the LLMs default to it.

Is this opinion on claude code or claude the model?

I'm NOT saying it is, but without regulatory agencies having a look or it being open source, this might be well working as intended, since Anthropic makes more money out of it.

Now that you point this out, since I started using Claude my RSI pain is virtually non-existent. There is so much boilerplate and repetitive work taken out when Claude can hit 90% of the mark.

Especially with very precise language. I've heard of people using speech to text to use it which opens up all sorts of accessibility windows.

Sorry to hear that and whilst it wasn't my original goal to serve such a use case I wonder if being able to interact with Claude Code via voice will help you? On MacOS it uses free defaults for TTS and ASR but you can BYOK to other providors. https://github.com/robdmac/talkito

Are you using dictation for text entry