Even if you're pro-AI, Anthropic themselves have said that using AI when learning means you don't develop skills
You need those skills later to debug the LLM's output or prompt it more effectively to fix something. These students are gonna struggle as soon as they have to do something harder.
For clarity on the current situation…
I did a PoC for some card scanning software a few years ago... Not something I want to sink the time into making GOOD. hehe
once I get closer to that, I may ping ya. Really appreciate the information and insights!
Another follow up - how do you do card identification for scanning?
I'm not sold on integrating card scanning with what I'm building; BUT - seems a feature people like and would tie into having images with the collection.
I'll dig into the TCGPlayer APIs more when I get to the pricing needs. I expect to pull from a few sources and mush it all together.
Quick glance looked like the affiliate program is still going, but there's some reading to make sure desired access is there.
My plan is cached. I'd like a 4x a day refresh. I plan to have some marketplace integration, so more accurate the better
I'm not looking to be /the/ current price, but relatively accurate with trending data.
I thought tcgplayer didn't let people into the API anymore? I will revisit that assumption
One thing I'm going for with v4 is more current pricing; do you mind sharing how you get pricing data?
V4 is underway; my "what to collect' has shifted. I'm using PostgreSQL for for relational-data. The JSONB and GIN look like solid solutions for expected queries structures.
Cosmos is really good for many things, but not all things. I found the edges of the system. :)
Didn't have any issues. They're all identified as unique cards in the scryfall data. The 'oracle'-uniqueness isn't a factor for how I work with cards. I also built out my own 'all printings' store for displaying all of the same cards. I would link to things, but azure is breaking my site for a bit.
V3 is stored as 'raw scryfall' and the way i approached viewing/tracking the cards was by set and card - I always knew what I was looking for.
Card entry is on the card display itself, so it always knows exactly what's being added.
Cost & Experience as drivers.
It's been a "my tool" thing, and I wanted to pull the cost down. I use Cosmos a lot at work; expanding on the tooling/knowledge for that has been really helpful.
MSSQL was used on v1; gets costly.
Never used PostgreSQL before, and has no work tie in.
I don't scan, which simplifies a bit. Data entry speed is paramount. All keyboard, no confirmations. I don't have to wait for it to finish the add before I can add the next card.
Backed by a cosmos database. C# backend, react front.
Version 4 will have a new data architecture.
I got an ask about AI and Quality - I smashed my face on the keyboard for a while to get quinngil.com/2026/02/24/a...
I made it; you can suffer it.
4 straight years of price increases... BLEH
You can get a Qobuz plan for less than I paid in 2018.
I switch from Amazon Music to Qobuz [https://www.qobuz.com] earlier this year. Mostly 'cause f-amazon. I saw they were going to charge $20 more; a 10% increase... made me wonder about the past...
2018-22: 149
2023: 159
2024: 169
2025: 199
2026: 219
Biggest concern for me with generated code - is it written how I want it written?
Pretty much; does it look like I did it myself?
If not; it's not good enough. Then I seek ways to enforce the code be the way I want it to be.
I think my next step is custom analyzer rules for my really picky bits.
I'm really hopeful (and everything indicates it will) enable much better code generation from UI to DB for my site.
I'm even using Claude Code to identify gaps in the memory files as I work through the layers to create the memory files.
My current AI efforts with Claude Code is building the memory files. I've gotten working with Claude pretty tuned in with JUST a root CLAUDE.md.
But there are a few things that are almost ALWAYS missed. The memory files should fix those.
Opus 4.6... Haven't gotten to test it across the app yet. Soon, I have some new features. But - it's doing the small things better, so I have high hopes for the bigger things.
It failed. It couldn't follow examples across the full application. But I could do a single layer... mostly... with review. And correction. And fixing.
But... faster than doing it all myself.
Opus 4.5 was a HUGE leap. It did better. Still review and correct, but less effort required.
I was trying to change my flow to maximize AI use... and it failed me. I switched to having Claude do the code I was thinking of. It did... OK. I normally had to give it an example to follow. It followed examples REALLY well.
I learned how to use it better, and I tried big again...
I've rebuilt my Mtg Collection Tracking site using AI. Minus a few core libraries that I copy/paste around.
I started with Claude Code in June of last year... it was... a struggle. It was good. Best around for sure. But I struggled to use it for big things. It never generated what I wanted.
It's behaving more like a dev I've tasked to do something than obviously an AI doing a thing.
While working on other things, turned a foreach into a Select. Which... is nicer. And exists elsewhere in the code. It's helping make things consistent; which is ... GLORIOUS. I love consistent code.
Am I happy it's making changes without ask... ehhhh...
Been off for a while; Prob gonna post some AI thoughts.
I have a lot.
Jumping back on because @claudecode.bsky.social with Opus 4.6 is... a substantial jump in the little things that make the code that much better.
In this morning's coaching session with an iOS developer, primitive obsession was getting in our way. Specifically, an array. I learned from @softwareascraft.com to take an array and turn it into a type you can control. "If you don't do it now, you'll end up doing it later after you regret it."
I had a test I couldn't see why it was failing. I paused TCR to work on it... TCR wasn't a problem though. The problem was I took too big of step. I didn't do the simplest thing. I knew where I wanted to go and I went there... and it made it harder.
Simplest possible. Sometimes we need reminders.
Continuing to use #TCR on this rewrite.
With the script to beep at me and automate the frequent commits... it's just #TDD. There's a few things a little different. Having done #TDD a lot - this is nothing new. It's strong support for what I want to be doing anyway.
I'm lovin' the small steps. Once the TCR challenges to normal refactoring flows have solutions... things'll go smoother/faster.