Human-Like Thought Processes
I'm constantly amazed at how similar advanced LLMs are to human brains in how they learn and solve problems.
Before LLMs, one of the big wins of Agile Development was "Working software over comprehensive documentation."
Anyone who's learned anything novel knows it's easier to learn by getting your hands on it than reading about it. Yes, studying does help and there are plenty of times when you simply need to memorize things, but truly understanding how something works often only happens once you can hold it, twist it, pull it, and bop it.
Bad Advice: Plan Mode
I first started using ChatGPT ~4-ish years ago to write little snippets of code for me; easy, repetitive things that probably qualified more as boilerplate that I could have just made snippets for, were I that organized.
With Copilot and then Claude, that became me asking it to write larger bits of code and then hours of asking it to change things because I didn't like how it did this or that.
Then came Plan Mode. The idea of getting better results by planning it all out. This was hard for me, because I like to get something working first and iterate on it. Ergonomics matter a lot to me, and I like to squish and prune code until it's neat and trim and does what it's supposed to do in as few lines as possible with maximum readability. Because I know I will have to read it again in the future, and if I don't, I will curse past-Sean.
So I started using Plan Mode. A lot.
AI Slop
This is when I first started to feel like I was using AI to produce slop.
This approach resulted in huge amounts of code that was difficult to read and understand. But also, I often didn't like the code.
It is SO MUCH HARDER to iterate on a huge amount of bad code than it is to grow a small bit of working code into a final implementation. During this period, I think I was a much less productive developer than when I just wrote all the code myself.
Back to Basics
My approach has since changed. I don't use Plan Mode much anymore, if at all.
What I do instead is give Claude a single prompt that I write, by hand, describing what I think matters, covering as much as I can get out in 3-5 sentences. That approach seems to trigger a thought process in the LLM that I find leads to the results I want, and much faster than Plan Mode.
Basically, get Claude on the right path at the start. The whole picture, no detail, just the key elements, and let it reason from there.
The Results
It works. Given the same model, I like the code produced much better. It's far easier for me to read and follow. I reach "done" far, far sooner than when I used Plan Mode. I am happier with the results.
It turns out, humans were right all along. A little theory at first, then going hands-on, is the best approach to understanding. Not walls of text. Not long, complex plans with everything spelled out in great detail. Clay in hands, squishing, pulling, twisting, bopping.
Summary
Give Claude a prompt to implement a working implementation as a starting point and iterate from there. It beats Plan Mode every time.