Jonathan Geibel

Disney Animation Tech Alum, AI Startup exec, founder, coder, product and engineering leader. Dad, tinkerer, pizza chef.

[Jonathan Geibel @jongeibel]

I've spent a couple of days having Astra just go do live product testing in the app and it just sort of worked which is cool (albeit slow). I then wondered how good Astra would be at building out a more complete Maestro end-to-end testing harness (basically, click through every aspect of the app on the iOS simulator) and that was even better. I normally don't like doing full e2e tests because they can be fragile and hard to maintain, but if Astra can keep them fresh and build them so fast maybe this is the right way to do it.

[Jonathan Geibel @jongeibel]

@kevinkern Interesting - that seems worth trying. "Go use the app like a customer would, specifically test things we've changed recently, let me know what doesn't work" -- a far more intelligent end-to-end test

[Jonathan Geibel @jongeibel]

My flow (today) is fable does all planning & implementation, astra reviews, paste review back into fable to decide what to fix, paste what fable fixed back into the astra review thread to check again, repeat until nothing major is found. At this moment this seems to provide the best results for me

[Jonathan Geibel @jongeibel]

@chrisstanchak Best 90s demo ever. Cool use of AI

[Jonathan Geibel @jongeibel]

I’ll keep hunting. Fable’s attention to detail is just next level. At the moment codex is relegated to code reviews for me. Traditionally I’ve found codex to be better at backend work but I’m also finding it hitting context window limits faster than fable and then degrading heavily after it compresses

[Jonathan Geibel @jongeibel]

Without a doubt, Fable > Sol by a pretty wide margin

[Jonathan Geibel @jongeibel]

Codex has really jumped quite ahead of Claude Code over the last few months in many dimensions. A bit surprised. Hoping CC gets their groove back on

[Jonathan Geibel @jongeibel]

Claude code with opus 4.7 has gotten super verbose with its explanations recently. Not a fan of that in AIs or humans

[Jonathan Geibel @jongeibel]

@nikitabier @BillyM2k I wish grok could automatically flag highly questionable posts, kind of like an early community notes indicator. Nothing definitive, just a hint on the most egregious posts while CN does its thing. Maybe I could hit a button such that once it’s resolved I get a notification

[Jonathan Geibel @jongeibel]

@ingoa_dev Codex and I had some words together

[Jonathan Geibel @jongeibel]

Maybe it's just me, but getting MCP servers to work in Claude Code is so much easier than the Codex GUI. Seems like such a simple and critical thing to somehow not get right on the Codex side

[Jonathan Geibel @jongeibel]

@mcuban Curious how you managed the herniations. Surgery? Yogo/stretching? Other less obvious solutions?

[Jonathan Geibel @jongeibel]

@tasedan @floraai Curious what your using it for

[Jonathan Geibel @jongeibel]

@haider1 They desperately need a coding environment

[Jonathan Geibel @jongeibel]

@kevinkern I hope that continues. Fallbacks continue to be one of the worst behaviors of coding agents. If that gets solved that’s actually a big step forward

[Jonathan Geibel @jongeibel]

I'm always amused when claude code says things like "or is this a good stopping point for today?" Like, does it need to be somewhere? What are its working hours? Or is it suggesting I should be doing something else soon?

[Jonathan Geibel @jongeibel]

@stevemur Pretty sure they’re still going to be on his ass

[Jonathan Geibel @jongeibel]

@kevinkern I think one type of skill that might last long term are skills provided by companies to make sure LLMs are using the best practices to accomplish a task. So, centrally maintained and used by many

[Jonathan Geibel @jongeibel]

@brian_lovin Seems like we are all inventing the same current best-practice Claude skills at the same time. Seems like someone should be able to get a starter kit with a host of best-practices skills as a baseline. Is there an effort to share and reuse skills?