I’m Jacob Bennett, a Staff Engineer at Medium working on AI enablement across our platform. I write about software engineering, product strategy, and decision-making—with a particular focus on how to build better systems and help teams make sense of complexity. Most of what’s here comes from lessons learned building at scale, and I try to share the thinking behind the work along with the conclusions.
We should all be using our brains, not offloading our thinking to AI
Crawlers should distinguish intent or be blocked by default (*ahem* Google)
No more prompting Claude. You write loops that prompt Claude. We’ve moved up another layer.
Run multiple review agents in parallel, fix the current bottleneck
Google’s post announcing OKF 0.1 as a formalization of ideas floating around.
Nvidia solved a large security gap for agent skills.
Some things are more nuanced than “definitely good” or “definitely bad”. AI falls into that bucket. It’s naive to say AI is definitely one way or the other. But you should always be pro-humanity.
An AI agent skill that searches across multiple platforms and ranks results by real-world engagement signals.
goose was built by Block and is now managed under the Linux foundation. It powers Stripe's "minions".
Notes from a moderately-successful book club I ran for Designing Data-Intensive Applications.
"What model does it use?" is the wrong question. Most of the differences you feel between AI tools come from the harness, not the model.
A collection of leaked prompts for AI tools (Cursor, Claude Code, Dia, Perplexity, etc). Idk how they keep this up.