Open Season

One of the primary reasons I've been a little light on my posting schedule recently is my focus on coding agents and a few useful AI tools.  I mentioned this before (or maybe elsewhere on LinkedIn - this is the danger of cultivating so many social media presences.  I confuse myself), but I finally broke down and decided to pony up for a Claude subscription to see if it was worth all of the hullabaloo.

It's certainly useful, but it's certainly not the second coming of sliced bread.  However, around the same time that I started using Claude more frequently, I also found out about Google Stitch.  I decided to use both in tandem to rewrite Slowpoke (again).  Of all the AI tools I've encountered so far, Stitch has by far been the most impressive.  In only 8 1-paragraph prompts, the first starting with this and performing the lion's share of the work:

I'm interested in building a personal financial app with a responsive design.

generated views like this:

I've asked for designs for a few other projects as well, and they look fairly polished, even if they're all a bit same-y (but what is the web if not based on lemming-centric design?)

I was also able to take the view above, feed it to Claude Code, and have it generate a skeleton page in Astro as a starting point for iteration.  I hooked it up to the sandbox version of a FinTech company called Teller to pull down mock transactions with the expectation that I'll be able to fetch my own data soon. 

Because I don't have the financial bandwidth to Tokenmaxx like all you Silicon Bros do, I have to watch my budget.  I've come close to hitting Claude's 5-hour limits on my Pro plan, but have yet to face the shame of outright rejection.

However, I was intrigued enough by the coding agent's capabilities that I decided to explore a couple of other options.  The first was the gemini cli.  It worked but seemed to be less capable than Claude Code (granted, the sample trial run size was 1, but Gemini in general seems to be a bit less polished than Claude - maybe because I rely solely on its free model).  I thus resigned myself to continue using Claude while being cognizant of its token limits and realizing I will just be less productive than everyone else programming these days.

But, on Sunday, I found an article on an alternative agent named OpenCode that I'd actually stumbled across a few weeks before but forgot to research.  Because I'd already fought the Easter Bunny to a standstill while it tried to steal my Cadbury eggs and had nothing better to do for the day, I decided to wade into the waters of expected disappointment and give OpenCode a try.  And...

...it did not disappoint!  Sure, it faces the same problems with client-side code generation that Claude Code faces, but in the now nearly 3 solid days I've been using it, it seems to be equally adept, if not more so, at generating useful code and following along in design conversations.  Its interface, especially when firing it up via `opencode web`, seems better than the razmattazing one that is Claude Code, and it appears to be much faster.

And, it's free!  Sure, I'm probably helping to sell state secrets to China (I'm using a model called MiniMax M2.5) rather than destabilizing my own citizenry, and, sure, I'm still contributing to the Earth's 2-pack-a-day habit to hollow out our environment.  But it's free!

The site does have a big asterisk saying that the models it's using are free for now (it appears that when the current batch is no longer free, they'll provide lab rats like me other versions to train on), but even if I have to start paying, the costs for the same usage on Claude are 50% or 75% cheaper (I'll have a better idea after using it for a month, but that's the estimate so far).  

It's enticing enough, despite the usual LLM flaws, that I've put off other crucial tasks (really, how does one need to brush one's teeth?) to spend time with it.  I'm sure the luster will wear off soon, and I've nearly sworn at it at least a dozen times (which is far fewer than Gemini), but I can see it being a mainstay in my toolkit.  

And, what's more, to follow up on the rant I posted on LinkedIn about how tools like Claude "democratize" coding by forcing people to pay to use them, this actually comes closer to doing so simply by shadow-supporing an authoritarian regime (but aren't they all, now?) rather than making you pay through the nose.  I'm still just sniffing around the Kool-Aid rather than drinking it, but now some of the tools do have the faint whiff of an enticing adulterated grape-flavored drink.

Until next time, my human and robot friends.