Claude can't estimate effort until it does the effort

I've just spent exactly one hour and six minutes with Claude Code on the next epic I'm building for this application. That's as long as it took a mix of two models and two effort levels to take an architecture document I wrote earlier today (time spent: 1 hour, 3 minutes), analyze it, read the codebase, figure out the architecture, identify the gotchas and things I missed, and generate seven full-fledged implementation specs adequate for Claude Sonnet to work from.

Before doing any of that, I asked Claude Fable 5 to read the document and estimate the effort and cost, keeping in mind I have a $100 credit from Anthropic specifically for evaluating the Fable model. Claude thought it would take 6 hours and cost me up to $100. Well, I already told you the whole thing took about an hour, and it looks like I only used $29 of the promo credit¹. I pointed this out when we finished and it replied "thanks for the feedback!" And then charged me for 1k tokens. Derp.

I'm still giddy about this. Just over two hours to spec out 8 full features. Wow.

It's going to take me a while to read through all the documents. Of course, that's why I had Claude write them separately: so that I can deal with each one as I deal with the feature in question. Given how quickly I've been able to build out features so far, all without runing over my subscription limits, I have a feeling I'll be able to release this entire epic next week. And this epic will be very visible to end users. I can hardly wait.