James Reagan

AI engineer / 🇺🇸 US Marine Corps veteran

[James Reagan @jpreagan]

@businessbarista @tenex_labs the word content is fascinating because it tells you nothing about what's being made and everything about how it's being distributed

[James Reagan @jpreagan]

@kentcdodds inevitable either ship openclaw or basically build out the idea piece by piece

[James Reagan @jpreagan]

@atmoio honestly it's plain for anyone to see

[James Reagan @jpreagan]

@intellectronica real talk

[James Reagan @jpreagan]

@hambleton_j @jturntdev tmux

[James Reagan @jpreagan]

@intellectronica @warpdotdev ok now i actually would like to take a look with that move

[James Reagan @jpreagan]

@GregorySchier debian stable just (always) works

[James Reagan @jpreagan]

@vincent_koc curious whether openclaw intentionally avoids transcript rewind/edit/resend, or if it just hasn't been built yet

i rely on this constantly in other agents jump back to an earlier turn, edit, and rerun from there

i'd be interested in adding it for webchat/tui, but wanted to check first in case there's a design constraint i'm missing

[James Reagan @jpreagan]

@MrAhmadAwais this depicts dollars spent and not tokens used, right?

[James Reagan @jpreagan]

plan mode was a mistake

every time claude code forces me in to plan mode i have to answer a series of questions that are often not helpful

i have to read and answer a number of questions seems like tedious busy work and wastes my cognitive bandwidth

then my reward is this wall of slop "plan" drives me mental to have to iterate in this manner with the clanker to get the plan right

worst. ux. ever.

why can't every session just be a conversation?

when i realize context is missing i force it into the context window either through tools or running bash commands with !

plan mode will just waste your time and often doesn't get it right anyway

while gpt 5.5 is my preferred model for coding, you can use opus 4.7 with pi, for example, skip the plan mode and prove that this isn't a problem with the model but the harness

overall the experience is so much cleaner

there is so much noise nowadays, i prefer to work with minimalist tools that push in my direction. i want to steer, don't harness-handle me.

when it is apparent to me the model has the right and sufficient context in its window, i've defined a feedback loop so it can iterate, and spelled out what done looks like then i'm fine checking back in hours later if need be but this is done by feel and my confidence in the model's response in our session

but plan mode ain't it

[James Reagan @jpreagan]

@webology @burkeholland oh yea i remember that one, man, looks like they're still active too

[James Reagan @jpreagan]

@JamesMontemagno i'm so excited to build native apps too. never done this but it feels so right!

[James Reagan @jpreagan]

@gabrielchua curious if you personally use this or yolo for non-work stuff

[James Reagan @jpreagan]

@tszzl some people with their lids ajar never heard of tailscale

[James Reagan @jpreagan]

@VictorTaelin > it would be 1m+ tokens worth

longer than lord of the rings? curious what domain we're talking about here.

[James Reagan @jpreagan]

With the right context and some periodic clarification or correction, I've yet to witness any coding problem that opus 4.7 or gpt-5.5 cannot handle.

I can hardly imagine what this next generation of models will bring.

[James Reagan @jpreagan]

@burkeholland "are you still talking to your computer?"

[James Reagan @jpreagan]

people of pi get unfiltered goblin mode

56 cents well spent

[James Reagan @jpreagan]

@burkeholland I'm coming to the same conclusion, working more not less.

And not only that but feel like it's got this other aspect to it that it is so much fun and highly addictive.

I get done with work and I just work on... open source like I don't want to stop!