Bart Proost

Dad, dev @anyscalecompute. Side projects and

[Bart Proost @BartProost]

Through our partnership with Google, I get to repost this cool tweet

[Google Cloud Tech @GoogleCloudTech]

Through our partnership with @anyscalecompute, Ray Serve LLM on Google Kubernetes Engine now offers 5x higher throughput and 8x lower latency for distributed inference!

Read more about scaling inference without the bottlenecks → https://t.co/9xV5vZ6pLE

[Bart Proost @BartProost]

@thedayisntgray @tobi @wintercarver Channels have a purpose and a purpose has an associated workflow, I can imagine. Keeps memory workflow-scoped so a gotcha in 1 domain doesn't pollute or bloat another.

[Bart Proost @BartProost]

One of my fav Ghostty features: cmd + k actually clears the terminal. Genuinely one of the most calming things in my workflow.

[Bart Proost @BartProost]

"Where to level up: all 12 commits landed on 5 of 7 days - no commits on Mar 21-22 (Fri-Sat). The 4-day streak is strong but shorter than the team's 68-day run. Consider queuing smaller PRs for gap days to maintain momentum." @garrytan's Gstack wants me working 68 day weeks 😅

[Bart Proost @BartProost]

AI is very good at doing the thing you already decided was important.
Much worse at telling you you’re looking at the wrong thing.

[Bart Proost @BartProost]

@thsottiaux Plan mode. Is there one? I couldn’t find it. I switched over fully at first but now only use it for complex concentrated tasks after Codex got very lost making lots of assumptions

[Bart Proost @BartProost]

@mattpocockuk I don’t think it’s bad, but it can be careless. Doesn’t really care if it introduced a duplicate helper or atomic component. I let an agent go ham on tracing associated symbols, finding references for everything it plans to build and use git to understand intent. Helps somewhat

[Bart Proost @BartProost]

@jamespotter Thanks for not dying though

[Bart Proost @BartProost]

@leerob Gpt 5 mini? There’s something special about that model’s orchestration skills

[Bart Proost @BartProost]

@mattpocockuk I think a big part of why plan mode is so powerful is the codebase research. Since it’s not actually writing any code you can gather context for a ticket while you work on other tasks, even without git worktrees

[Bart Proost @BartProost]

@joshpuckett That is beautiful. Love that the parent decreases in size a bit when you open a submenu. Great detail

[Bart Proost @BartProost]

@robj3d3 Since Claude doesn’t write code during this step, you can prep tickets in parallel even without using clunky git worktrees.

I (or Claude) wrote a Cursor extension to automate this. You can find it at https://t.co/KSeX7Esn1G

[Bart Proost @BartProost]

@robj3d3 Copy the output, then use /clear and start planning + implementing.

I’ve been using this at work for a while now, and it dramatically changes how may PR comments I get.

Another benefit about this workflow 👇

[Bart Proost @BartProost]

@jarrodwatts I (or actually Claude) wrote a Cursor extension to do this automatically and prep using tickets from Lineae and Jira. It makes the process less cumbersome. For who’s interested: https://t.co/BdylzETDB4

[Bart Proost @BartProost]

@jarrodwatts At my day job this has made a huge difference in the amount of comments PRs get.

Extra benefit: even without git worktrees you can prep tickets in parallel using this workflow, since it doesn’t write code 👇