John Webber

lunch time!

@thsottiaux Remote control

What are we obviously not getting right with Codex?

It’s very difficult to a reference doc(s) that a model can reliably turn into a working product. Simple stuff is much easier but add any complexity and it’s a real challenge. But I think it’s a good idea because it: - eliminates a lot of security concerns: supply chain attacks don’t really exist when your model writes all/most of the code itself - you can tweak the app to your exact needs very easily - new features are just new reference docs, you decide which ones you want to “install” (I’m really excited about this part and working on a POC a lot lately) Then instead of having to be a great programmer to write great programs, you have to be a great architect and writer.

I vibe coded myself a very helpful project+ agent management tool. But I kept adding and removing random ideas and features and eventually it was spaghetti code. So I had gpt5.5 write a bunch of md files as references of all my features and core functionality…. …then I ran a ralph loop for 7hrs to rebuild it all from scratch. Now I have really clean and well organized code but a lot of core features that don’t quite work right. if I iteratively vibe code an app, it works because I’m testing every step along the way but my code’s a mess. If the code is clean the features need a ton of work.. Gotta figure out the middle ground here.

Never tell Claude to write docs for your repo. They go stale too fast. If you really need docs, make it a scheduled task so at least they stay up to date.

time to go home!

@icanvardar opened my account 17 years ago and right now i have 16 followers. that's hard to do.

devs with 0 to 250 followers are often the ones building the most interesting stuff

@ProductHunt Remote Claude Code (that works)

pitch your product in 5 words or less 👇

@edhumbling @acrogenesis @Cloudflare blog.cloudflare.com/ai-labyrinth/

@johnwebber @acrogenesis @Cloudflare John , what does the labyrinth do ?

Starting today, agents can now be Cloudflare customers. They can create a Cloudflare account, start a paid subscription, register a domain, and get back an API token to deploy code right away. cfl.re/4sY0Uxn

@Shpigford Just email directly from your Gmail. ….or mail them a letter. The “software” all has a smell to it.

i'm slowly coming around to cold email as a sales strategy 🫣 that being said, WHY IS ALL OF THE SOFTWARE IN THIS SPACE UNIVERSALLY TERRIBLE?!?!?!?! i've tried a dozen options and every. single. one. of. them. is the worst software i've ever used.

I mainly do e-commerce work so my clients want to be scraped and indexed by ai bots so they get recommended more in the future. (One of my companies actually bought a smaller competitor for the sole reason they outranked us heavily in ChatGPT searches for our products) We’re already seeing a huge shift in traffic from Google search over to ai recommendations. To get an AI recommendation you have to be scraped and included in their training data, right? Or mentioned a bunch on Reddit….. So why would a site owner want to block AI bots? Don’t they just lose out on a huge chunk of future traffic from a source they are now invisible to?

@philfreo @johnwebber @acrogenesis @Cloudflare Why doesn't it make sense? You mean they would want to let their content be crawled by ai labs for free?

I’ve used it for a few client projects to give me some great starting points. After that ui(dot)sh is helpful for different ideas. Claude Design did feel a tad slow vs just using Claude Code for iterations. Hopefully it inspires people to make something better and we get a big wave of design tools like it. Prompt > Images 2.0 for quick ideas > Claude Code for html preview That’s a pretty simple flow someone could chain together into a nice interface

Who is actually using Claude Design? Anyone?

Yeah I agree that longer context has gotten better quality in recent updates. I have still found I get much better results when doing fresh sessions though. I view my agent conversations as somewhat ephemeral though, I want my context + memory setup to make it easy for fresh agents to find what they need quickly.

@johnwebber @embirico Yeah that is an issue although imo increasingly less so. What I do to counteract is i make sure when i want it to run long, my latest message is a good one, which in turn guides it better and keeps it from going off the deep end. It sees the latest msg again each time it compacts

Do you prefer queuing or steering in Codex? What's your use case for each?

@curiousharish Use Claude/codex and the AWS CLI. You’ll still have to learn the service names but you should be able to provision everything without ever logging in to AWS.

Someone please build a wrapper on AWS! Most confusing interface ever.

Claude just used a word I’ve never heard before…”munging”?? (My spellcheck even tried to change it just now)

Cloudflare’s default for new sites is to block AI bots. And they have the labrynth option which is specifically designed to waste tokens. I’d say that’s pretty adversarial against AI. Now they allow an AI agent to create a new account. Again, I’m all for this new attitude, I love CF and I’ve been using them for a long time and their new AI features are awesome. But it does really feel like a pivot over the last few months

@johnwebber @Cloudflare No pivot. They still block ai training which was the fight they were putting some months ago

@posedscaredcity @embirico Have you compared this vs a new sessions for each task? I’ve found it far more effective to have a new session for each task. Context and memory help a ton but I feel like if I let it go as far as compaction I start to lose quality of responses and work

@johnwebber @embirico Yeah I maintain one context window per project or line of investigation, compaction plus memory has been really good. I steer not to give it new tasks but to modify the existing one based on how it’s doing ie “spend less time here and explore this other thing more”

@posedscaredcity @embirico In the same context window? So if it’s doing task A for days, you’d jump in and steer it by giving it an entirely new task to do right inside that same context window?

@johnwebber @embirico Hmm I leave it running for days every so often I check in and steer it. I think it’s very useful in that case. It’s like being able to send a slack msg to a coworker