rob

Building practical Codex tools and AI workflows.

[rob @robdocx]

Long Codex session?

I built Clean Handoff: a local-first Codex plugin that creates a trustworthy project checkpoint, verifies the live repo, and helps a fresh session continue without re-explaining everything.

Looking for early testers:
https://t.co/lrpsbhcHox

[rob @robdocx]

@DFdavidfontaine The correction has to pay off right away. If it feels like labeling data for the product, people will stop doing it. It should just feel like they’re making their own tool work better.

[rob @robdocx]

@kevinkern 8GB per VM is a tough tradeoff for everyday iteration. I could see this being more useful as a clean final verification pass when you need to know the app works outside your local setup.

[rob @robdocx]

@thebitstack @X Appreciate it. I’m still figuring out which workflows are actually worth packaging as plugins. Curious what you end up building with Codex.

[rob @robdocx]

Looking to connect people on @X

if you’re into

- vibe coding
- UI/UX
- Claude Code
- Codex, Cursor
- build in public

say Hi or drop what you’re working on looking to follow active ones 👋

[rob @robdocx]

@guoqingfeng6 @X Nice. Claude Code gets more interesting once the product is real enough that you’re making tradeoffs instead of just generating screens. What are you building?

[rob @robdocx]

5.6 feels less like an incremental upgrade to me and more like the first model that consistently understands the whole job. For GPT-6, the leap probably won’t just be “smarter.” Realistically, it’ll be autonomy, reliability, memory, and the ability to operate for hours without drifting.

[rob @robdocx]

@emilkowalski Useful pattern. I’d want the skill to keep the source material attached so you can see what was copied, inferred or added later. Otherwise it gets hard to update once the original thinking changes.

[rob @robdocx]

@rileybrown That boundary is rough because a skill is only useful if it follows the work. I’d rather define it once and choose where to run it than rebuild the same setup for every chat mode.

[rob @robdocx]

The more I use coding agents, the less interested I am in saving every bit of context.

Old decisions are worse than missing context because they still look trustworthy. I’d rather give a fresh session less information that still matches the project.

[rob @robdocx]

@rezoundous The downside of making the best model everyone’s default: eventually everyone tries to use it at the same time 😂

[rob @robdocx]

@kirakissaten Using AI for basic tasks is not inherently pathetic. We’ve always built tools to reduce friction. The real question is whether people are using that saved time to think more deeply, create more, and learn faster, or simply outsourcing every thought.

[rob @robdocx]

Long Codex sessions keep ending the same way for me. The code moves forward, but the context gets fuzzy.

I’m building Clean Handoff so a fresh session can pick up the project without me explaining it all again.

[rob @robdocx]

@kr0der That’s a good use of it. Are you having Codex handle the slide structure too or just the assets and export? Getting the whole thing ready to airdrop is the interesting part.

[rob @robdocx]

@YJBee @X Nice. I’d be curious how you’re structuring yours. I’m separating the project goal, decisions, constraints and live repo checks so a fresh session can tell what is still trustworthy.

[rob @robdocx]

@niek_olthof @X That makes sense. I’m curious how you keep the saved context useful without letting old information pile up. I’ve been treating verification and cleanup as part of the handoff instead of trusting every saved note forever.

[rob @robdocx]

@plainionist I tune them whenever I see a repeated failure mode, not on a fixed schedule. One-off misses go in notes. Repeated misses become an imperative constraint, then I test it in a fresh session. Otherwise instruction files turn into a junk drawer.

[rob @robdocx]

@ChatGPTapp @dkundel The browser handoff is surprisingly useful when a task moves from reasoning into real UI work. Letting the agent handle tedious navigation while keeping the user at the publish or submit boundary feels like the right balance.

[rob @robdocx]

@OpenAIDevs The inline accept, edit, or reject loop is the important part here. Keeping the proposed patch reviewable before it touches the branch makes Codex feel much more like a collaborator than a black box.

[rob @robdocx]

@Mattkleonard78 @X Very cool. WebGL transform feedback is a wild foundation for an internet OS. What’s the next milestone?

[rob @robdocx]

@BrettKileeg @X Absolutely 🤝 What are you building right now?

[rob @robdocx]

@thebitstack @X Great learning project. Password generation gets interesting fast once entropy and UX collide. I’m building Clean Handoff, a Codex plugin for project checkpoints and session continuity.