Pinned Tweet
Most of my 30+ apps were built before coding agents became my default. Now I use them in real app development and document what actually changes: faster output, more parallel work, and new bottlenecks in review, testing, and shipping.
1
1
555
I made a little game about that split second when the puddings wobble and somehow stay up.
I just shipped it. 🍮
#gamedev
3
2
55
A useful way I’ve found to use scheduled tasks in Codex: retrying after a short wait.
If GitHub is temporarily unreachable or a site is under maintenance, I can ask Codex to try again in an hour and continue the task. I don’t have to remember to come back and restart the work myself.
24
It looks like Codex can now access information from ChatGPT. I haven’t caught up on the details yet, but this could be useful when asking Codex to process a batch of tasks based on things I’ve already discussed in ChatGPT.
I’d love to be able to do the same with Dot voice conversations, though that doesn’t seem possible yet.
32
Stick, swing, switch. Tap to swap feet and climb from window to window.
A little game I'm working on.
#GameDev #IndieGame
2
5
53
My biggest challenge with Dots voice is getting instructions to apply at the right scope. Sometimes they don't stick; other times, a request meant for one situation carries over into almost every exchange.
I also find myself repeating how I'd like it to respond. Am I missing a way to make those preferences stick across voice calls?
In my experience, voice also hallucinates more than text chat, which makes it harder to use reliably. But being able to move work forward through conversation is genuinely powerful. I'm really looking forward to improvements in instruction following and reliability.
41
dot's replies are quick and fluent, but I've also had it tell me a task was done when it hadn't actually done it, on multiple occasions.
Just now, it said it had watched a video and even gave me its impressions. The response came back so ridiculously fast that it was immediately obvious it couldn't possibly have watched it.
I use dot almost entirely through voice, and hallucinations seem more noticeable to me than when I use Astra directly in Codex. I haven't compared voice and text within dot, so I wonder if this is specific to the voice model. Has anyone else had a similar experience?
1
43
Met up with friends who I’d consider fairly plugged into tech, but no one brought up AI agents. I’d started to think everyone was vibe coding their own tools and systems. Maybe my corner of the internet is a lot smaller than it feels.
1
1
50
I’m working on a new game idea. In the meantime, here’s Oh No Cat, a cat-themed drawing puzzle game I released a few years ago.
32
I’ve started using dot to give Codex tasks by voice while my hands are busy. It’s handy when it works, but sometimes a request sounds acknowledged and never reaches Codex. I still find myself checking whether the work actually started.
1
1
50
I found that dot could read a local Codex thread on my Mac through an existing Codex cloud thread. I’m starting to experiment with using that context to talk through progress and product decisions with dot over voice.
33
What I like about using dot by voice: my eyes and hands stay on the actual work.
I can review a document or playtest a game while talking things through with AI. Some of the work I used to put on a second monitor now happens through voice.
24
I can now ask Codex to work on my Mac from my iPhone via text. That’s what I was hoping for. Voice requests still aren’t getting through in my setup, which is a shame. Hoping OpenAI gets that working too.
38
I’ve spent roughly 6+ hours using dot over the past two days, organizing admin work and discussing game ideas. Voice responses are fast, and it handles fairly complex conversations well. It’s been useful as a thinking partner.
But I’ve run into problems reading other threads and receiving generated images, and I was told recurring tasks couldn’t be set up. So far, I haven’t been able to delegate work as smoothly as the official docs suggest. I don’t know what’s causing it, but getting these integrations working reliably would make a big difference.
34
Voice chats with dot usually feel really natural, with a comfortable speaking pace. But sometimes it suddenly talks super fast. Is it matching how fast I’m speaking, or does its pace change on its own? Curious how this works.
39
My ChatGPT Pro account got 62,500 Codex credits, so I added remaining-credit tracking to the Codex Usage Meter I built for myself.
Now I can see weekly usage, credits, and rate-limit resets in one place.
1
1
117
I caught a glimpse of what SaaS without a UI might feel like.
I connected an AI agent to my bookkeeping software via MCP. Rule-based automation works for familiar transactions, but a new pattern across a batch still needs review and judgment.
I reviewed the entries with the agent, approved its proposed changes, and had it update the books. I’ve used this software for over a decade, but I’m no expert in its UI. Since I don’t use it every day, I spend time remembering how to do things. For this kind of work, the UI had been the friction.
1
26
I have Codex open too, and the messages from this call are updating there as we talk. This is pretty much the experience I’ve been hoping for: one conversation that carries across voice and Codex.
1
1
10


