Obligatory XKCD:

https://xkcd.com/303/

The main problem to me is that you do not always know if you are going to wait for 1 minute or 10. That can be very annoying. On good sessions though the back and forth is relatively stable, but too short to really invest your brain in something else and too slow to just stare at your screen.

And context switching is tiring, so if I can I write specs for the next prompt / next task, otherwise I just surf the web. You can also review the code as it is edited to save time.

I carefully look over her shoulder to read all the thinking outcome. To ① know what the heck do we do, ② be able to interrupt her at any time she’s driving bonkers, and ③ make it easier to CR after everything is done.

Huh? It’s extremely dangerous to give an LLM tasks you could not accomplish yourself (because you won’t be able to validate the result.) But if you can write the code you probably should know how much time would it take.

So yeah, I always know how long am I to wait (and it’s never 10 mins, there is next to zero chance LLM is able to accomplish such a long task without hidden uncertainties, which will surely shoot her in the foot later.)

No it’s not dangerous if you know what you are doing. It has to read all the specs I wrote, understand the full codebase to make the right changes.

Review is not a problem as I know the codebase very well. (I wrote 80% of it, if not 98% depending on projects).

Admitedly my worse timings were on slow 4g network but still, Claude with high thinking, peak usage and a large codebase can be so slow.

But if you can write the code you probably should know how much time would it take.

I know how much time I would take to write it, yes.

I’ve always said that the basic concept of algebra (and indeed of all mathematics) is proportion!

This looks like a nice saying yes, but its not worth much more in this context.

It’s very obvious to me that when you open a codebase (or just started claude) you will take a lot of time to do the most basic feature, whereas when you know the codebase (or claude has many relevant files in context) you will take much less time to do something that is actually more complex than the first task.

Sometimes I look at my side projects for any small wins, other times I catch up on slack. Mostly talk to friend on WhatsApp (this gives me extra energy to continue with my work day at the cost of my wife thinking I dont work at all.

I had one project this week that Claude, on Sonnet 4.6 high, took 4 days to complete - with no input after the initial prompt. I was surprised it took that long but it did. Not a general purpose project by any means, and it had a formal spec, a conformance test suite and several reference implementations to work from so it really didn’t need me in the loop.

What do I do when it’s doing its thing? Have a few more sessions doing other things!

Or, increasingly, go outside and enjoy the fresh air. And construct something physical, not virtual

I tend to do a combination of [1] follow along with it (though it can start to feel crazy after some time because of how quickly it goes), [2] work on other parts of the project, [3] research or do external marketing work, and [4] step away from the computer to be present in other areas of my life.

I’ve noticed the most gains in quality from Opus 4.6 (and 4.7 is a bit shaky at times —> either being spot on or quite sloppy, and I did read that you need to prompt it differently than you would for say 4.6). In Tidewave, I’ve tended toward using 4.7 (except for yesterday it seemed to behave strangely) for fixing specific a11y compliance and 4.6 for less specific coding problems (like working on new features or reviewing things). So, despite the cost, Opus 4.6 has been a positive shift in working with an LLM (kind of like having someone to do the day laboring for you while you architect/manage the work).