After reading about it for a bit, it seems like Claude Code sprinkles in system reminders periodically to encourage the agent to use the skills jannesklaas.github.io/ai/2025/07/2...

I imagine the real trick behind skills is whatever they're doing to get the agent to make use of the progressive disclosure efficiently. Maybe tuning? Maybe they periodically remind the agent that skills are available?

Even prior to skills, I experimented with introducing progressive disclosure to my context md files - having it be a "directory", pointing to other md files that had more details about specific topics, but I find that the agent just ignores those links.

Rational Funk with Dave King. youtu.be/XOTZttTVEqY?...

And that's how you get the gig.

strongerbyscience is great for no-nonsense, evidence-backed nutrition advice. Some of it angles towards "getting ripped" but it is generally grounded in science and approaches it with a great 80/20 mentality. You might want to start from this guide: www.strongerbyscience.com/diet/

They don't :(

So the best way to get access to gpt-5 for this purpose is to set up a script that burns through $50 of tokens to get you to tier 2. That's absolutely bonkers! Just let me preload my account with $50 of credit to get access to the higher tier!

30K TPM in tier 1 basically prevents you from being able to use a coding assistant. And there's no way to pay your way to a higher tier, you have to get there by organically spending $50

I develop a coding assistant and OpenAI feels really hostile towards developers. Their sdk is a mess (does not adhere to jsonschema, is poorly documented).

Recently I tried to run some experiments on gpt-5 and immediately ran into their rate limits / tiering system.

Open source tools should just let you define a regex for finding context files. (Like `autoContext` in magenta github.com/dlants/magen...)

- I added authentication via anthropic's max and pro plans, which lets you use tokens the same way as claude code (thanks to the opencode folks for their open-source implementation for how to do this). This is much cheaper relative to using the pay-per-token API for heavier users.

- I renamed "@compact" to "@fork", and also changed it from a forced tool use to just a regular tool. This means the agent can think before using it, improving the quality of the summary, and also this now takes advantage of the conversation cache, making these tool calls a lot faster and cheaper

i tested it, it works

🎉 github.com/dlants/magen...

I can now have multiple ghostty windows open, one per tab, and use a command shortcut to quickly switch between them :tada:

Now I just need to figure out how to make my ghostty sessions persistent, and do some tweaking around how tabs are numbered, and I'll basically be there.

this uses hammerspoon to launch a ghostty terminal. There it runs a script that uses hammerspoon's cli to query the open windows and tabs (using the stuff that ghostty exposes for a11y). It then feeds it into fzf, and when fzf returns a selection, uses hs again to switch to that window/tab.

After a night of hackery, I was able to replicate something workable using #hammerspoon and ghostty's accessibility features! github.com/ghostty-org/...