Rasty Turek

You are the sum of notifications you allow.

Here is list of good and bad behaviors with a practical model: - 2-4 original posts/day - 5-15 high-quality niche replies/day - 1-3 quote posts/day if you add real analysis - 0-2 reposts/day

@michalillich No. I added those for clear visual and mental scoring. Otherwise it’s super hard to keep track of what matters and how much

@synopsi is it really adding hard-coded points? That's how I programmed things 25 years ago. I would expect heavy machine learning being used today...

Ran bunch on analysis on the algo to identify what the algorithm is penalizing and promoting. There is no smoking gun like if you do this it works. Instead the algorithm really looks for audience alignment.

So the optimal post: -> clear niche/topic -> strong match to target user history -> survives all hard filters -> predicts dwell/reply/share/follow -> avoids not-interested/block/mute/report/no-dwell -> not crowded by your own other posts -> passes visibility/safety checks -> earns real positive actions -> strengthens future retrieval and ranking

@0xSerious Kontext is a competitor to hasp gethasp.com/comparison/

@SeveTom @synopsi @hasp_dev Look into kontext.security

It is clear to me that we will have to rebuild the whole security suite for agents: - Password manager - Firewall - Antivirus I got first covered (@hasp_dev), working on the second and thinking about changing mdgrok into the 3rd.

@filip_ornstein Yes. One stable assistant that holds some of the most important context and delegates to specialized agents. Exactly that

That’s a good idea, but the subscription model still means spinning up CLI processes, so it gets expensive fast. I’m thinking more about the layer above implementation: how to let an AI product/analyst role make more dynamic decisions across whole epics instead of me. The goal-based flow is already pretty good. But I still catch myself pretending I need to stay in control too often.

The gain we are making with agents tooling overshadows the gains with models. I haven’t thought about a model for quite some time. Even qwen 3.6 does fine with well set up harness and proper tools.

@filip_ornstein Yeah same. Using top models for some tasks. The biggest unlock for me is identifying use cases for different models and proxying them through those. I am constantly thinking about new kind of harness where I can run different family of tasks through different models.

Tools bring the context, do the boring lookup work, guide the reasoning flow, and plug in skills. That’s more context again. A frontier model with a million-token window won’t beat an iterative, focused harness. I think about this every day while running five CLIs at once. The next quality jump won’t come from models alone. We should be looking for the next layer above them. Still using 5.5 though ;)

@tom_herudek I assume it’s the update that is canning copied data. Mdgrok has very little of unique content.

@synopsi Any idea why? Maybe too much generated content?

@seraleev That’s weird. I got new app released in March with trial enabled. Issue updates weekly. No problem.

Makes sense. I have 3 apps where I haven’t released updates since last year (and since the end of January 2026, App Review no longer allows updates with the trial toggle enabled). I also don’t have remote configs set up, so I’ll make time in the next few days to update my apps. Thanks for taking the time to help me become better.

I guess Google really really really doesn't like mdgrok.com Essentially sent the traffic to 0. Site still gets around 3k daily users, but this stings.

@jtaby Yeah the tooling is improving rapidly.

I'm implementing a major project in Codex. I've been in a single thread for over a week. Context management and compaction in codex app has been remarkable given where we were months ago. Haven't had the need to clear context and start over.

@SeveTom @hasp_dev Yup. Looks directionally similar to what I have in mind for the firewall

@synopsi @hasp_dev Look into keycard.ai

@erichstark @hasp_dev I use sets of skills and recently added mobbin's MCP but haven't used it yet. Also using ui.sh and impeccable.style

@synopsi @hasp_dev Do you use some tool for design preparation? It is done before plan or after? When I tried for new app it seems to be inconsistent at first.

Reading proficiency of black students in California in the year before Newsom took office: 32% In the most recent year: 32% So, a 66% increase in spending led to a 0% increase in reading proficiency.

@dshaw000 @kimmaicutler Yup.

@synopsi @kimmaicutler New York City does?

I applaud the governor and also am annoyed that this omits relevant context that state governments statutorily must balance their budgets and can’t run deficits like the federal government.

@dshaw000 @kimmaicutler Both state mandate balanced budget. Claiming that they done it is equally disingenuous.

@synopsi @kimmaicutler I don’t believe those are the same