I skipped customer interviews for years. So I went on @asmartbear's Smart Bear Live and Jason Cohen walked me through his iterative hypothesis method, live, on my own interview plan. It was eye opening youtu.be/1QFJFCDnWGo

I'm sad about the openai resets. the models aren't that good, and i'm only using opus 5.5 right now. a month ago that seemed impossible, and i'd have loved these almost daily resets. they really need to fix the models. I think openai winning is better for us consumers.

Have to say fable 5.1 better than astra at debugging issues with chatgpt app. Chatgpt's @ chrome connection stopped working in @heliumbrowser today. It worked fine yesterday. Astra told me to restart Helium and the chatgpt app (already did both). Then it told me to reinstall the "browser control plugin." I found no such plugin. So I asked fable. Fable 5.1 went through the logs and found it: Helium reads chrome's NativeMessagingHosts folder to find chatgpt's connection file. macosa 27's App Data protection blocks chatgpt from that folder. Today's chatgpt update deleted the old file and couldn't write a new one. Fix was simple: Go to Privacy & Security → Files & Folders → give ChatGPT Google Chrome access, restart Helium. Now it works again.

Learned a lot about customer development from Jason. It was so good! Thanks @asmartbear

Edit your landing page like a Google Doc, then hand every change to Claude Code in one prompt. @jessethanley said it's cheap to rebuild, so I had Opus 5.5 build my own. Open sourced it so you can skip that part. github.com/risonsimon/agent-…

Opus 5.5 is giving me an Opus 4.5 level jump in experience. So good. It’s also much cheaper than I expected. I tried it with Cursor, and it’s cheaper than Grok for completing tasks. I’m also really liking gpt 6 sol because it’s so efficient. I’ve been using it all day, and it’s barely moved the usage on my $20 plan. With Astra, I’d be done for the week.

Do chatgpt chats now consume codex credits? I had about 50% usage remaining, but I did a couple of chatgpt chats on the web (not Work). I came back to 0% usage left. I don't understand. I haven't created any cloud tasks. If it wasn't chatgpt, what could it be? @thsottiaux

One of the biggest problems with routing models is that they can break the cache. But cursor trained their classifier to account for cache misses. So if I have a context heavy conversation, does that mean it’ll never switch to a smarter, more expensive model?

Saw this launch & I thought I'd try self hosting it. It was way easier than I thought. One message to codex & took 9 mins. It deployed to cloudflare free tier & set up search console integration with computer use To set it up yourself, send this to Codex/ChatGPT app: First login to cloudflare & google cloud console on chrome, connect codex to chrome. And install wrangler cli for cloudflare. "can you set this up for me? github.com/every-app/open-se…? clone the repo and set it up and if you need anything from me, let me know" You need to give it dataforseo and openrouter api key, but that's it. It's possible because @bensenescu made the docs so good and made it work on cloudflare stack.

I avoided worktrees for months because they seemed unnecessarily complex. Turns out I was wrong. Now I can test 3 different product flows at once without touching my main codebase. Here's how I got worktrees working with @convex in codex. (works with claude code too) I started by asking codex to research the setup. Told it to check both codex and convex docs and come up with a plan. Codex created a script for setting up the worktree, and copying over the relevant env vars. First hiccup: codex was using outdated convex cli commands. I had to manually copy the new commands from convex's changelog and ask codex to use that. Without this, it created overly complicated scripts. Second hiccup: I didn't realize I had to manually select the new environment for each worktree. Once I figured that out, the setup script ran automatically. For convex specifically, I needed to set a specific node version for it to run locally. But that was it. I also set up portless by @ctatedev and Vercel Labs so that each worktree has its own website. It works beautifully. Just need to create a branch in the worktree, and portless will create a URL with that branch name. Another reason worktrees worked for me is using @better_auth . Because each worktree can run fully isolated, each one keeping its own login state in the local convex. It’s beautiful. Now I can test different product concepts without touching my main deployment. Everything runs locally. No extra costs. Worth the setup time.

Just had a holy shit moment with codex computer use. I asked it to archive all newsletters from my Missive inbox. Took some time but got through everything without errors. I've tried multiple AI inbox tools before. None come close. This uses the actual app on my mac. Way more powerful than I expected.

I don't get the Opus 4.7 hate. Ya, it's more expensive with the new tokenizer. But performance wise? It's solid. I've only used it in @cursor_ai till now and it's been fine. Maybe it's just a claude code issue.

Anyone got this update yet?

I'm adding 🟡 to Codex thread titles just to track which ones I'm still working on I don't love archive. It hides threads completely. I just want active ones on top and archived ones visible below, separated. How is everyone else managing this?

What happened to @GroqInc?

Who do I need to tag in @convex to get a response?

They got me.

2016: what if Google builds your idea 2026: what if Anthropic builds your idea

Hey @waynesutton, I upgraded to Convex starter plan and all my referral bonuses disappeared. That shouldn't happen right? Can you help?

Why does Codex cli have /compact but the mac app doesn't? I don't want to start a new message at 90% context window. Right now I'm opening codex cli just for this.