Condensing Twenty Years of Wisdom in One Markdown — by Konstantin Gredeskoul, 08/19/2026<br>// AI Condensing Twenty Years of Wisdom in One Markdown<br>August 18, 2026 Konstantin Gredeskoul Comments<br>Estimated Read Time: 50 Minutes<br>#ai#agents#claude-code#postgresql#context-engineering#skills#database#schema-design#indexes#vacuum#pgvector#connection-pooling#productivity<br>I Have, Once Again, Become a Manager, and I Kinda Love It.

I’ve been building at a genuinely unreasonable pace lately. Open source on one monitor, private commercial work on another, and somewhere in the middle of that a startup I am trying to bootstrap into existence before my savings account files a formal complaint.

And here’s the thing nobody warned me about: at this point I feel very much like the manager of a ten-to-twenty person elite engineering team. Not metaphorically. Structurally. The team never sleeps, never argues about tabs, never takes a sabbatical to “find itself” in Lisbon, and never once asks me whether we’re doing OKRs this quarter. It just sits there, idling at full capacity, waiting.

Waiting for me.

I am the bottleneck. Me. The specifications are the bottleneck. The acceptance criteria are the bottleneck. “Did you actually click the button and see if the thing does the thing” is the bottleneck. I have become the single-threaded process in an otherwise embarrassingly parallel system, which, if you have ever profiled anything in your life, is exactly the outcome you spend your entire career trying to avoid.

So I am on a mission to automate myself out of that position as aggressively as possible. Along the way a few things fell out of the process that I think are genuinely useful to other people, so here they are.

The Great Skills Gold Rush of 2026

Meanwhile, in the broader ecosystem: every other week somebody publishes a repo of “Andrej Karpathy’s skills” - they are not, in fact, his - and collects twenty thousand GitHub stars before lunch.

And we are all encouraged, loudly and continuously, to curl | bash a pile of instruction files from a stranger’s repository into the context of an agent that has read access to our entire filesystem and write access to our entire codebase. Why? Because it has a lot of stars. That’s the whole security model. Stars.

We spent fifteen years building supply-chain scanning, SBOMs, dependency pinning, Dependabot, provenance attestation, and reproducible builds, and then collectively decided that a markdown file which literally reprograms the thing writing our code deserves less scrutiny than a transitive npm dependency. Bold. Genuinely bold.

I am not going to install your skills. I am going to write my own, because I actually know things, and because the only person whose taste I have to trust in this arrangement is somebody I have been arguing with for twenty years: me.

The Inspiration: Kun Chen, and Doing It the Old Way

My latest round of workflow surgery was directly inspired by Kun Chen’s video, “L8 Principal Building a Full Stack App with Agentic Engineering”. I cannot overstate how worth watching this is. Go watch it. I’ll wait. The post is long, you’ll need the stamina.

This guy is young, and yet he attacks problems the old-school way: you don’t like how something works? Write your own. And write he did. lavish-axi alone is priceless - a local-first editor for the HTML artifacts your agent produces, so that the review loop stops being “take a screenshot, paste it, write four paragraphs describing what’s wrong with the padding” and starts being “click on the broken thing and tell it.” HTML is the new markdown, annotations are the new code review, and somebody just went and built it instead of tweeting about how somebody should build it.

That is the correct instinct, and it is in desperately short supply right now.

About That 500x

Let me put a number on the pace, and then let me immediately qualify it into the ground.

In terms of pure lines of code produced - code that works, that follows the best practices I set, that I did not trust and therefore manually tested to death on top of the 96% coverage the test suite already gave me - I am currently something like 500x more productive than during some of the most productive stretches of my career, including pairing with the founding engineering team at Wanelo.

Five hundred. Times.

Now sit with the second-order consequence for a second, because it is not a victory lap: nobody has time to review this much code. Not me. Not you. Not the very good engineer you’re about to hire. The review bandwidth of a human being did not go up 500x. It went up approximately 0x, and arguably down, because now the diffs are enormous and uniformly plausible-looking, which is the single worst property a diff can have.

Which brings me to the actual point of this post.

The Only Leverage That Scales Is Taste, Written Down

If you cannot review the output, you have to constrain the input. That’s it. That’s the whole trick. You move your judgment upstream,...