Today I finished one of the least visible, but most important parts of my AI engineering workspace project.

Command Guard

I keep running into this while building.

There s always another feature that sounds useful on its own.

Another setting.
Another metric.
Another automation.
Another smart layer.

But at some point, the product starts asking the user to understand the system instead of the system understanding the user.

I've noticed a shift in how I work over the past year, and I'm not 100% sure it's a good one.

A year ago, I'd use AI to speed up execution like writing boilerplate code, drafting an email or summarizing a doc. The thinking was still mine. Now, more and more often, I catch myself asking AI to decide things too: which architecture to pick, how to frame a pricing page, even how to respond to a tricky customer email. The line between "helping me think" and "thinking for me" has gotten blurry.

Here's what worries me a bit: the moments where I learn the most are usually the moments where I struggle first. If AI removes the struggle, does it also remove the learning? Or is that just an old-fashioned way of looking at it - the same way people worried calculators would ruin math skills, and mostly they didn't?

Do you think AI has made you better at thinking, or just faster at producing? And any moment where relying on AI backfired because you skipped the "struggle" step?

Trust is especially important when an app can access location, motion sensors or emergency contacts.

However, transparency is not always as simple as publishing everything. Sharing source code and data flows can help independent review, while exposing detailed detection logic or security controls could also make a safety system easier to bypass.

Which information would actually help you decide whether to trust a personal safety app?

Would you look for a clear privacy model, permission explanations, selected public source code, independent security reviews, data-retention controls, or something else?

The situation that happened:

I see how some creators on Subtsack are monetizing their content (they have a section for subscribers who can pay to access articles, videos, and private chat).

At this point, all of the AI coding assistants are in the same neighborhood. Decent at "advanced autocomplete", OK at code generation sometimes, and most are somewhere in the process of incorporating code context mechanisms. But what's next? Agentic behavior? Something else?
My pet prediction is that we will see the emergence of a new programming language that's designed for use with AI and can be translated to a variety of popular languages. (Or if we're cursed, just javascript )

Most productivity tools optimize for completeness: more activity, more metrics, more detail. But a useful update is not a transcript of everything someone did.

It should explain what changed, why it mattered, and what happens next. My coding agent is not able to do this that well because of the context gap

Every credit system I've used bills on the API call, not the outcome. So when a model returns something plausible and wrong, the user pays, then pays again for the version they actually wanted. You've priced your own failure rate and handed the bill to the person who noticed it.

The counter I keep hearing is that inference costs money either way. True, and beside the point. If a generation is unusable because we routed it to the wrong model, that's our decision, not their prompt.

Two things fall out of eating it yourself, both good. You have to define unusable narrowly enough that it can't be gamed, ours is a same-prompt regenerate inside a short window and nothing else. And you start caring a lot about routing, because a bad route now costs you directly instead of quietly costing someone else.

Curious who's actually doing this and where it broke for you. Most pricing pages I read still bill per attempt and call it usage.

A founder coach who surveyed hundreds of builders said something that stuck with me: every single person she interviewed used the word "lonely."

That really stopped me.

I've just made my free tier worse for a group of people who did nothing wrong, and I can't work out what I owe them.

Background, anonymised as best I can. Until this week you could use the core function of the thing I build without an account at all turn up, get a result, look at it. You only had to identify yourself if you wanted to keep the output. This week I put a free account in front of it. Same allowance, same features, no card, but you now sign up before you get anything.

The reasons were dull and real. I couldn't attribute anything to anyone. I couldn't tell a returning person from a new one, so I had no idea whether people were coming back. And an anonymous endpoint that costs me money on every call was an abuse problem I was going to have to solve eventually regardless.

What I didn't think through is what to say about it. Three things I'm stuck on: