![]() | |
Thomas et al.,
Claude 4.7 Opus and I are drinking buddies. We meet at VSCode tavern where we collaborate over code for Stormy the Stingray floorbot. While over the years I've been acquainted with about a dozen programming languages, I'm nowhere near a 'programmer'. He wrote for me an exploration SLAM mapping program, taking into consideration works by our ROS SIG colleagues. I observe and report on Stormy's behavior, Claude suggests parameter tweaks to improve it. I wanted to have warning lights and sound to improve Stormy's safety. Claude drew a circuit diagram in ASCII, specified components, and suggested GPIO pins on the Raspberry Pi. When the circuit was assembled, he wrote a python ROS2 node that used X-Box buttons to toggle headlights, flash orange LEDs, and synchronize Klingon alert sounds. It monitors NAV2 states to announce "Danger Will Robinson" when he gets into trouble. As my ideas changed, Claude suggested alternatives with pros and cons, adapted and modified accordingly.
He's not perfect. He's given me conflicting wiring directions but clarifies and is contrite when I point it out. Sometimes his advice is counterproductive or he misunderstands my intentions, but we work it out.
One BIG caution: keep your chats short and focused or they can get expensive. I've blown through a month's AI budget in a weekend. Remember the ENTIRE thread gets processed with each query. That helps greatly with continuity, but gets expensive if you keep adding new topics. Make each chat about one particular issue and stay focused. To maintain continuity between chats I add a context.txt file explaining my hardware, OS environment, and file organization at the beginning of each new chat.
At the end of the day I ask Claude to in PLAIN TEXT summarize what we accomplished which I paste into my "...Experience" log. I also have him make and give me the entire path of his memory file for the episode such as /memories/session/2026-05-30-explore-session.md so he'll know where we left off without replaying the entire dialog.
Earlier versions of Claude, even 4.6 were not nearly as good. Sometimes loosing track of the environment, context, previous instructions,.... Giving bad advice based on wrong versions of hardware, software, or context. Leading to unaccustomed profanities. That rarely happens with 4.7.
Vibe coding with Claude Opus 4.7 is like working with all of y'all with your multiple areas of expertise in the room advising me, but without the debates. Except Claude can type way faster and more accurately than I can! He can analyze a request and modify 80 lines of code in the time I can take a sip of my Nespresso Stormio caramel latte.
But I'm not willing to give him too much 'agency'. I don't want to hear him explain "...but when you said you really wished you had that NVIDIA computer with the super GPU on Amazon that it was ok to...."
James H Phelan "Nihil est sine ratione cur potius sit quam non sit" Leibniz
--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/94C7A979-BA15-4A52-8793-AAD4A4EA6C5F%40gmail.com.
--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/9ccc0c95-b81d-470b-bb7a-cbb208b6935fn%40googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/34b339d0-da24-4d19-97c3-a1f71e49ddea%40hal-pc.org.