Weekly Changelog #2
Originally published on Substack.
Foreword
I was surprised to see a recent YouTube video promoting the relatively old /grill-me skill by Matt Pocock. At 295k GitHub stars, it clearly has continued value as a skill even as models have continued to progress.
When I first heard of that skill, I instantly wasn’t really into it, and assumed it was a fad like many micro-optimizations that become solved upstream by model competency or harness optimizations. But since it has stuck around I wanted to understand why I felt such an instinctive rejection to it.
I think what it boils down to for me is that it becomes way too easy, borderline inevitable, for long term usage of this skill to erode one’s default sense of intellectual rigor. That is, I think asking questions is just as important of a behavioral muscle, if not more, than answering them. To question well requires understanding, and to answer well does too. In job interviews, I’m constantly judging based on the questions the interviewer/org has chosen to ask me, and I can’t tell you how many companies don’t realize how poorly this reflects on them. Better yet, show me an engineer who is willing to ask earnest questions when asked to act immorally, or break the law - and I’ll bet you they’re pretty good at their craft. Well-reasoned questioning isn’t the only signal, but it’s a great one.
And so I’m not going to pretend like this concept broadly needs background - everything from Plato to scientific peer review revolves around collective question-and-answering as a way to learn, validate, and discover. But when we place the model eternally in the asker’s seat, we remove a lot of the learning and growth potential of a human in the loop, even if the code churned out is deemed satisfactory.
Being able to critically analyze and ask probing questions helps me learn as a code reviewer, and (hopefully) leads my colleagues to better net results, both of which are part a key reinforcement loop that eventually begets a skilled, seasoned, and properly jaded engineer. The friction of being wrong, having to correct a mistake, or update one’s out of date understanding is not time wasted on human folly, but exactly how we fortify our talents and critical thinking.
At least for the moment, my context of a decade plus engineering still tends to beat a model’s hypothetical questions born in the fever dream of whatever the closest AGENT.md file has soothingly injected into its smooth, smooth brain. The skill isn’t inherently bad, but like all new technologies, I think planful and conscious adoption is key until deeper understanding is achieved. And I guess I’m bearish on that take going forward for as long as progress dominates the conversation and long term effects on learning, mental health, and psychosis are conspicuously absent in the broader discourse. When it comes to human learning, we need to be careful.
On that cheery note, below are a few articles and recordings I’ve been digesting that have been informing my thoughts about how well my relationship with AI actually serves me:
"Rather than being 'too connected,' I think it's dangerous to make decisions because you're being guided," says Kojima (via an edited machine translation). "I'm not denying technology but, for example, AI on my smartphone introduces me to all kinds of things. I don't like that either.
"That's because I think 'coincidences' are necessary in human life. When you wake up in the morning and go to school or work, you might happen across a coffee shop. Or you might meet someone by chance. I think it's a series of such events, and the accumulation of choices, that creates your life.NYTimes: Opinion | Is It a Relationship if No One’s on the Other Side?
Engineers don’t care about the inner life. They don’t really believe it exists. It’s not useful to them. It’s not necessary to them for their mentality of what they’re trying to build. And so I think that there is a danger in taking an engineering product that does not have an inner life and using it as your best friend or your interlocutor, except when in very confining circumstances where you’re very clear that you’re getting information.
Releases
lowly-writing-framework
An emergent composition framework skill for writing issues, PRs, comments.
My lowly-writing-framework v4.3.0 went out this week with some bug fixes and tweaks around reviewing pull requests. The skill should now better understand some of the nuances between an informal review and a literal “Review” in the GitHub feature sense, as well as using emoji reactions to indicate sentiment and softening negation comments/replies to humans.
lowly-writing-framework-plugin v1.1.0 is merely a dependency for the embedded skill to catch up to v4.3.0.
foulflock.com
Foul Flock cross-references crowdsourced ALPR camera locations against state camera-siting statutes.
lowlydba/foul-flock 2026-10-17 has the first tagged release for foulflock.com. The build had been broken since Overture Maps Foundation changed its taxonomy in the September release, which is now fixed.
Additionally, the map now renders in different layouts for tablet/phone sizes so the site isn’t only usable on Desktop. PMTiles were also introduced and the cold load went from 45MB to 9.4MB pre-compression.
While Texas, Florida, and Connecticut don’t have siting laws, some updated research indicated there are various guidelines (i.e. highways right-of-way) that could be mapped still. The certainty for this data is a lot lower, since in some cases one would need to know the actual installation date of an ALPR at a given site to determine is legality. That being said, 10% state coverage is progress!
Tofu Garnish
🌿 Tofu and Terraform outputs, plated as static sites. A GitHub Action.
lowlydba/tofu-garnish v1.3.1 had a minor UX improvement so that now the copy button keeps a fixed width on click, instead of temporarily re-sizing post click.