Posts

Here’s my latest AI setup and a few things that have changed about how I work.

As of this writing

My core application is ChatGPT in the Codex section. I use Astra 6 on Low reasoning as my starting point for almost all new threads. I start in a project folder when I know what I’m working on, and remove the folder when I don’t.

I use bypass permissions and computer use. I use Field Theory for voice to text and context stacking. Lately, I’ve also been asking the model to keep Linear up to date for more complex projects, so I can pick work back up without returning to a long-running thread and spending tokens rebuilding context.

Acceptable clutter

A long list of threads is a common end-of-day state for me. I used to worry that it was too much clutter, but as I rely more on the model to create threads, I’m starting to think of them as “working threads” and “visibility threads.”

Working threads are my coordination with the model. Visibility threads give me some observability into the model’s delegation of work to subagents.

I don’t worry too much about how many there are anymore. I often right-click a thread, copy the deeplink, and give that to a fresh thread when I want to pick up where I left off.

“Please rename my threads” is also a very useful command.

Less plan mode

I’m no longer using the dedicated plan mode in any harness. I’ve found that asking the model to “repeat back what it understands I’m trying to achieve” is often sufficient for me to agree with or correct its thinking.

Plan mode fell away for me after the release of models like Astra, Fable, and GPT-5.6 Sol.

Codex beats Claude

I’m finding Codex substantially better at doing useful work, especially in how efficiently it uses tokens.

In my experience, Fable 5.1 trails Astra by a reasonably wide margin across most of my work: design, engineering, coordination, and computer use. I expect this to shift as new models arrive.

I still use Fable periodically as a subagent to weigh in on decisions that I’ve asked Astra to orchestrate.

“Subagent review”

I say this a lot after a large context dump and an initial discussion with a model about a project.

Sometimes I specify, “Please ask Fable 5’s opinion.” In my setup, this requires being logged into Claude Code, and I sometimes need to remind Astra that it has access.

It works nicely. Subagents, even when I’m only using Codex, do a great job of refining project scope, reducing complexity, and catching potential issues.

“Be token efficient”

You can tell Astra this directly, or say, “Assign work to the appropriate subagent reasoning level.”

I’ve found that it does a good job of not chewing through my tokens.

Session resets

Another useful part of my setup is banked usage resets. In Codex, I find these under Settings › Usage and billing › Usage limit resets.

If I run out of usage, I can use a banked reset to get back to 100%. They do expire, so I keep an eye on them.