pull down to refresh

100 sats \ 2 replies \ @optimism 1h \ parent \ on: Stacker Saloon

Cool.

  1. I'd edit/review in vscode anyway so it's all in one place (if I were in vim or emacs, I'd probably use the clis)

Yeah I'm a vimmer so that's a bad assumption on my part - if you anyway use an IDE then that makes some sense. I just review on forgejo now that I've pipelined claude code. I no longer run it in a terminal unless something goes terribly wrong, then i just restore the session.

  1. I like model/mode switching

Why?

  1. Cursor's plan mode is my favorite AI thing right now

You get that in the cli tools too, and it often just gets triggered when going over the codebase. I no longer implicitly use it... if it wants to plan it can plan. Sometimes I ask it to comment on an issue with a plan instead of implementing it.

My workflow is:

  1. Write issue
  2. Tag issue with either of these:

  1. Work on something else or go shitpost on SN
  2. When I feel like it, look at notifications on isolated forgejo or my planning board that also gets fed by issues with bot/* tags. Review some PRs, merge what can be merged, write more issues.
  3. I run out of credits before everything is done all the time so I just reprioritize on the kanban and then I move on.
write
preview
100 sats \ 1 reply \ @k00b 1h
Why?

I like mode switching because I don't like it writing code without me understanding all the decisions being made. Modes are like different guard rails that help me predict what happens after I prompt.

I like model switching because I'm curious about the strengths/weaknesses of the models in my work/workflow. I'm using opus nearly all the time, but when a new model drops I switch to see if it'd be worth switching.

I just review on forgejo now that I've pipelined claude code. I no longer run it in a terminal unless something goes terribly wrong, then i just restore the session.

I'll end up here at some point. My AI workflow is close to my pre-AI workflow.

reply
100 sats \ 0 replies \ @optimism 1h

About a quarter of the time when I just "oneshot" an issue into it I get a PR where I'm like "oh shit I forgot to spec something". Those are the second most costly.

The most costly is when I think "oh shit I didn't spec it" and then I look and I did spec it and I'm like must be I'm typing runglish so I just close the PR, prefix the line in the spec with "IMPORTANT: ", make it bold, and requeue it. And then 99.99% of the time all is good.

Your approach is more precise than mine, i'd do the same when I'd be coding anything other than tools with the bots, but I won't. Not yet... because too many mistakes, also with Opus.

My AI workflow is close to my pre-AI workflow.

Mine is fully segregated. My pre-AI work workflow has AI review before human review now.

reply