• I Ported Kubernetes to the Browser: Sam Rose generated almost 100,000 lines of TypeScript with LLMs to build webernetes, a partial Kubernetes port that runs entirely client-side: pod scheduling, cluster DNS, and a browser-based container runtime. His verification process is more interesting than the port itself. He reviewed the generated code line by line against the Kubernetes Go source and ran hundreds of integration tests against both webernetes and a real k3s cluster to prove behavioral parity. The weekly token-cost and lines-of-code charts show what a two-month AI-assisted build actually costs.

• ua-tracer: What Does a User Agent Actually Do?: Paul Kinlan built a tiny site that mints a unique, secret-scoped URL on every page load so it can report whether a visiting bot parsed your CSS, followed your manifest, or executed your JavaScript. The early results differ by crawler: GPTBot follows CSS but skips JS, OAI-SearchBot runs full JS, ChatGPT-User does neither, and ClaudeBot fetches everything but stops short of parsing the stylesheet. This tool makes the mechanics of an AI “reading” a page visible.

• Lore Architecture Docs: Lore is a transparent HTTP proxy between a coding agent (Claude Code, Codex, OpenCode, or Pi) and its upstream LLM provider. It captures conversations in three memory tiers: raw temporal storage, distilled observation logs, and curated long-term project knowledge. Its five-layer “gradient context manager” decides how much context to compress from the actual cost of cache reads and writes on each turn. That makes context management an explicit economic decision.

• WSL Container Is Now Available for Public Preview: Microsoft has built a native Linux container CLI (wslc) and matching API into WSL. Windows apps can now run Linux containers without putting Docker Desktop in the middle. The new default filesystem (virtiofs) doubles Windows file-access speed, and the preview already works with VS Code Dev Containers. I’ll try it on the Windows machines where I currently keep a separate Docker installation just for Linux containers.

• Fastbelt Version 0.1: TypeFox’s new language engineering framework, a Go successor to Xtext and Langium, has reached its first real release. It includes grammar tooling, automatic cross-reference resolution, parallel workspace builds, and core LSP features such as completion, go-to-definition, and hover. TypeFox claims up to 50 times the speed and one-twentieth the memory use of comparable frameworks on common workloads. Anyone who has built a custom language server will recognize how much boilerplate that could remove.

• Data-Oriented Design in Yuku’s Parser: Arshad Yaseen’s JavaScript/TypeScript parser, written in Zig, handles an 8MB file in about 19ms by storing its AST as flat arrays of u32 indices instead of a pointer tree. It needs no per-node malloc, traversal stays cache-friendly, and the whole tree frees in one call. Because each child is appended before its parent, a linear pass over the node array is already a post-order traversal. Yaseen also explains why Yuku applies the technique less rigidly than Zig’s compiler: Yuku’s AST is a public API that other tools must use.

• Cloudflare Workers and Hyperdrive with TanStack Start: Part 2 of Adam Rackis’s Cloudflare series covers two recurring database problems on Workers. Hyperdrive’s connection pooling reduces the cost of connecting on every invocation, while TanStack Start’s global request middleware creates a fresh “connection” per request to satisfy Workers’ rule against persistent I/O resources. He also pins the Worker’s placement region to the database region, a small configuration detail with a direct latency cost when missed.

The post AI Agents & Systems Engineering: 7 Bookmarks Worth Your Attention appeared first on Cristian Nebunu.

The post On Privacy, Design & Career: 4 Essays appeared first on Cristian Nebunu.

The Oh My Posh Visual Configurator is a drag-and-drop builder for your terminal prompt: arrange segments, preview in real time, and export the config for PowerShell, Bash, Zsh, or Fish. If you’ve ever wanted to prettify your prompt but bounced off hand-editing a JSON theme, this removes the friction entirely. Small, but the kind of thing that makes the terminal feel like yours.

MonoLisa is a coding font with a live tester page where you can preview it in real code, toggle every OpenType feature (ligatures, stylistic sets, the grade axis, alternate glyphs), and try it across JavaScript, CSS, Java, PHP, and Python. It’s a premium font, but the tester is free and genuinely useful for deciding whether a wider, more characterful mono is worth the switch. Readability is the whole game when you stare at code all day.

Marc Brooker’s Surprising Economics of Load-Balanced Systems is an older piece (2020) but exactly the kind of evergreen theory worth carrying around. The setup: c single-slot servers behind a load balancer, offered load scaling linearly with c to keep per-server utilization constant. What happens to latency as c grows? The intuitive answers (constant, or worse) are wrong: via the M/M/c queue and Erlang’s C formula, mean latency decreases, asymptotically approaching pure service time, and the percentiles follow. Bigger fleets get better latency at the same utilization, and most of the benefit arrives at modest c. One of the rare distributed-systems results where scale makes a problem easier. File it next to your capacity-planning instincts and recalibrate them.

The post Today’s recommendations: A visual builder for your shell prompt, a coding font tester, and the counterintuitive economics of load balancing appeared first on Cristian Nebunu.

• Starting in the Analog Domain: Stoyan Stefanov sees the same pattern in music composition and “vibe coding”: committing ideas to a computer too early makes the work take longer. He develops music through an idea, an instrument, and paper before opening notation software. A computer’s precision makes it a poor place for loose experiments and quick failures. His short essay makes a persuasive case for sketching before building.

• I Hated Writing Until I Learned There’s a Science to It: An engineering student who dreaded writing receives a paper covered in red ink from a graduate-student mentor. The marks turn out to be the ordinary shape of revision rather than a verdict on her competence. Treating writing as a skill with techniques, instead of an innate talent you either have or lack, offers a better model for anyone who thinks they are “just not a writer.”

• Don’t Let AI Write for You: Sebastian De Deyne shares a short passage from Alex Woods about the purpose of writing. The work increases the writer’s understanding before it helps the reader. In Woods’s analogy, letting an LLM do that work resembles paying somebody else to exercise: you receive an output but miss the strength that should come from the effort.

• What’s Wrong in My Thinking About Errors?: Om Malik follows up after two resilience-engineering researchers challenged his earlier distinction between forgivable human error and machine error. Their correction is precise: “human error” often closes an investigation before it finds the systemic cause. Anyone who has filed an incident under “operator error” and moved on will recognize the problem.

• Write Some Software, Give It Away for Free: This essay makes a quiet case for building software as self-exploration rather than as a route to a paycheck. Work made that way may avoid the user-hostile, value-extracting patterns that appear when every decision serves a financial result. The Hacker News discussion is divided between maintainers whose generosity attracted entitlement and those who experienced the opposite.

The post On Writing, Craft & Thinking Clearly: 5 Essays appeared first on Cristian Nebunu.

Speckyboy’s 20+ Best Free Minimal WordPress Themes is exactly what it says: a curated roundup of clean, typography-first, content-forward themes (McLuhan, Integer, Koji, Rams, and more), with a handy comparison table at the end. Nothing fancy, but a good bookmark if you or a client ever needs a distraction-free blog or portfolio theme without paying for it.

Jamie Marsland’s two Pootlepress pieces are a pair, so I’m keeping them together. In WordPress Block Themes Were Made for AI he rebuilds his wife’s yoga studio site with ChatGPT and comes away convinced that block themes’ simplified, semantic template-plus-content architecture is exactly what makes them AI-generable in a way classic themes never were. Its companion, Using ChatGPT as a WordPress Page Builder, is the step-by-step version of that claim, his argument being that ChatGPT might be the best WordPress page builder he’s used, purely because of the speed of creative iteration. Read together they’re a good snapshot of where the WordPress-plus-AI conversation actually is right now.

Ethan Hawksley’s JSON-LD Explained for Personal Websites is the most immediately actionable thing here for your own site. It’s a copy-paste guide to the structured-data nodes that actually move the needle (WebSite, Person, WebPage/ProfilePage, SoftwareApplication, BreadcrumbList, Blog, BlogPosting) with the real JSON-LD from his site for each, and clear notes on @id linking, sameAs for disambiguation, and the fact that crawlers merge nodes across pages by ID while single-page scrapers (LLMs) don’t. Increasingly it’s not just Google reading this; it’s the answer engines deciding who to cite. If your site’s JSON-LD is thin, this is a weekend win.

And DashboardStack rounds it out on the app side: a CLI that scaffolds a database, auth, CRUD, forms, and Stripe billing for Next.js, Nuxt, or TanStack Start, then deploys to Cloudflare Workers/D1/R2, deterministically, so you’re not burning AI tokens regenerating the same admin plumbing on every project. From the maker of Nuxt Charts, code you own outright. A nice “start with the dashboard you’d have built eventually” tool.

The post Today’s recommendations: Minimal WordPress themes, whether block themes were made for AI, ChatGPT as a page builder, JSON-LD for personal sites, and a dashboard-scaffolding CLI appeared first on Cristian Nebunu.

• The Slow Death of the Power User: This essay argues that two decades of platform design choices have taught a generation not to understand its own tools. Its examples include the absence of a user-accessible filesystem on iOS for more than a decade and Android’s gradual restrictions on sideloading and bootloader access. The author believes the effect has reached developers who rarely leave managed abstractions. The argument connects technical literacy with meaningful consent: a privacy policy is hard to evaluate when the underlying system is inaccessible.

• First 10k Month From SaaS, 4 Areas of Focus, and a Mental Model for Growing a New SaaS: Justin Butlion’s monthly update covers his bootstrapped SaaS portfolio crossing $10,000 in monthly revenue for the first time, mostly through Monday.com marketplace apps. His “four areas of focus” model limits him to four active projects and rejects work outside them. I am going to try the same constraint on my side projects.

• The House Is Valuable Because It Is the House: Ethan Marcotte reflects on redesigning his personal site after The New Yorker described the aesthetic sameness of LLM-generated websites, including beige backgrounds and rusty orange accents. He chose texture over forced novelty: hand-picked typefaces that mix old and new, a watercolor motif, and a relief map adapted from a mountain near where he grew up. His redesign asks a practical question that has become more urgent: does this site look distinctly mine?

The post On Power, Attention & Building What’s Actually Yours appeared first on Cristian Nebunu.

David Wells’ Multi-Agent Coding Without Worktree Chaos is the most useful workflow piece I’ve read in a while. He went all-in on git worktrees for multi-agent coding, then drifted back to a single branch, because worktrees defer collisions to merge time rather than removing them. His answer is coordination, not git ops: plan the work, interrogate the model for non-obvious questions, turn the plan into a directed task graph, then let a swarm pick up unblocked nodes while a coordinator (“ring leader”) loops every few minutes. Collisions are prevented by explicit file reservations and an agent mailbox for blockers, not by git status archaeology. The line that stuck: “opening more terminals is not a system.” This is structured parallelism, and it’s clearly where the leverage is.

Which makes Herdr the natural companion: “tmux for coding agents.” It’s a Rust binary (no Electron, no account, no telemetry) that runs your agents in real PTYs on any box, keeps them alive when you close the laptop, shows blocked/working/done/idle state at a glance, and lets you reattach over SSH from your phone. Real panes, a plugin marketplace, and a CLI + socket API so agents can orchestrate the multiplexer themselves. If you’re running the kind of swarm Wells describes, this is where it lives.

Devashish’s Two Qwen3 Models on One DGX Spark is the deep-systems entry for anyone chasing local inference. The goal: run Qwen3-Next-80B-FP8 for heavy work and Qwen3-4B for fast turns coresident on one Spark (119.67 GiB unified memory), behind a LiteLLM proxy on a single endpoint. It took a weekend, and the write-up is the “residency math” pain: gpu_memory_utilization is a snapshot against total card memory, not free memory; Qwen3-Next is mostly Mamba, so KV-cache demand is dominated by Mamba state alignment, not attention KV; and coresident processes race: they don’t negotiate. Niche, but if you’re eyeing a Spark to own your inference, this is the hard-won map.

Two Agent Skills close out the cluster, both of which land right on your workflow. Josh Pigford’s SEO Content Generation skill is a self-contained content engine: drop it in ~/.claude/skills/, and each run researches what to publish next, picks the content type, writes one deeply-researched on-brand piece, and registers it in a dedup ledger, built for 2026 search, optimized to rank in Google and get cited by ChatGPT/Perplexity. Given your own recommendations series, the “decide → research → write → verify → register” loop is worth studying even if you never run it. And unslop-text is the one I think you’ll actually want: a skill that strips the cues that make prose read as AI-generated (the em dash, “it’s not just X, it’s Y,” sycophantic openers, delve/tapestry diction, the “in conclusion” wrap-up), and warns against the newer over-corrected “trying not to sound like AI” register (forced fragments, fake typos, bolted-on swearing), which it treats as its own tell. It’s grounded in a Reddit analysis of what readers actually name as giveaways, ships a deterministic scanner for CI, and is honest that the real tells (uniform rhythm, empty fluency) are structural and need a human ear. This is essentially your HARDLOCK ruleset as a portable skill: required reading for how you edit The First Legion.

The post Today’s recommendations: Multi-agent coding without worktrees, a tmux for agents, two LLMs on one Spark, and two Agent Skills worth stealing appeared first on Cristian Nebunu.

• Now We’re Getting AI Fake News Complaining About How AI Fake News Is the Death of Real News: Joshua Benton traces a viral story about 47 Alabama weekly newspapers supposedly killed by an AI content mill. Its source was a site with no real staff, invented bylines, and fabricated named sources, all pointing back to a Finnish tech company. He started investigating after noticing that every other “Politics” story on the site undermined Taiwan’s position against China. The result is unsettling: fabricated reporting can target a niche so specific that almost nobody has the local knowledge to catch it.

• The Human Story of the Open Web: Jeffrey Zeldman writes about a preview screening of “Code for the People,” a short documentary on reclaiming the open web from a handful of “digital landlords.” His prescription is the same one he has followed since the 1990s: keep your content on a domain you control and prefer open-source tools over proprietary services that can change the rules overnight.

• The AI Resist List: Researchers including Karen Hao, author of Empire of AI, built this living global directory of organized opposition to harms from large-scale AI development. Entries are grouped by the part of the industry they target, including funding, data, labor, surveillance, and policy. Examples range from a Japanese union petition about algorithmic wage-setting to a Uruguayan campaign for data-center transparency and an Amazon employee letter about AI and climate goals. The directory shows who is organizing, where, and by which method.

The post Trust, Ownership & the Open Web: 3 Essays on What We’re Losing appeared first on Cristian Nebunu.

Bramus Van Damme’s Cranking View Transitions Up to 11 is his Beyond Tellerrand talk write-up (slides and a recording of the sister talk are linked). The premise is delightfully reckless: push the View Transition API past its documented limits: combine it with scroll-driven animations, trigger it automatically with a MutationObserver, even resurrect classic IE’s page transitions through this modern API. It’s the tinkerer’s tour of the feature, and coming from a Chrome DevRel engineer it’s the good kind of “here’s what happens when you ignore the spec’s guardrails.”

Una Kravets’ Modern CSS Theming is the one to actually implement. She combines three newly-Baseline features into a genuinely nice theming system: light-dark() as the macro foundation (it reads computed color-scheme, so it’s inherited and composable, not a media query), a per-shadow-layer trick to swap drop shadows in light mode for glowing borders in dark, contrast-color() for one-line accessible text on any background, and @container style() queries that branch into hue-tinted palettes instead of harsh black/white. An optional @function --elevation() cleans it up at the cost of narrower support. This separates page-level theming from per-card micro-theming in a way I hadn’t seen laid out so cleanly.

Manuel Matuzovic’s field-sizing for selects is a small, satisfying fix for an old annoyance: by default a <select> is as wide as its longest option, so a control showing “Utrecht” stays as wide as “Scheveningen.” field-sizing: content lets the width follow the selected option instead. One line, exactly the kind of platform detail worth filing away.

Sebastiano Guerriero’s Lab01 (the CodyHouse founder) is a gallery of “best-in-class” UI experiments: a dozen small, polished interface studies you can open and inspect. Pure frontend inspiration; good to keep in the tab you open when a layout needs a spark. And Saazio is a conversion-focused SaaS website template built on Next.js 16, React 19, Tailwind 4, TypeScript, and shadcn/ui, with the usual marketing pages ready to customize, a sensible starting point if you’re standing up a product site and don’t want to design the hero for the hundredth time. (Relevant to your own publishing-tool and SaaS projects.)

The post Today’s recommendations: View Transitions pushed past the spec, modern CSS theming, sizing selects to their content, UI experiments, and a shadcn SaaS template appeared first on Cristian Nebunu.

Addy Osmani’s The New Software Lifecycle pulls the ideas that matter out of a Google whitepaper he co-wrote. The framing I keep reusing: an agent is a model plus a harness, roughly 10% model and 90% harness (the rules, tools, sandboxes, orchestration, guardrails, observability), so when an agent does something dumb, debug the harness first, because most agent failures are configuration failures you can fix today. Context engineering (static vs dynamic) is a financial lever, not just a technical one; verification (tests plus output and trajectory evals) is the line between vibe coding and engineering; and the lifecycle compresses unevenly: implementation collapses to hours while specification and verification stay slow, so spec quality becomes the bottleneck. “Set the bar at the eval, not the demo” is the sentence to hand a skeptical lead.

Armin Ronacher’s The Coming Loop is the uneasy counterweight, and, full disclosure, his third appearance across this series, so I’m flagging it (more on that in the dedup note; you may want to space him out). He’s watching harness-level loops (a queue, a machine that attempts work, a harness that decides whether it’s actually done and keeps it alive past “I’m done”) take over agentic engineering, and he’s honest that they already work astonishingly well for porting, performance exploration, security scanning, anything that transforms existing code or produces short-shelf-life artifacts. His worry is lasting code you care about: loops amplify the model’s instinct to add local defenses instead of making bad states impossible, and the deeper fear is dependency: codebases produced, reviewed, and maintained by loops that assume machine participation, that you can no longer understand without the machine. And you can’t fully opt out, because attackers and competitors will loop against you. It pairs naturally with the Osmani piece: same future, opposite temperament.

Brandur Leach’s The Minimum Viable Unit of Saleable Software is the SaaS-economics read, and a genuinely useful mental model given your own product ambitions. Starting from the viral “we replaced $400/mo Jira with a Claude-built tracker” anecdote, he does the actual math: cheap isn’t zero, LLM-built software still costs feedback loops and maintenance labor, and the Jira clone takes ~37 months to break even while a $25k/mo Salesforce is closer to a real build decision. He proposes a “zone of viability” (enough novelty that a rebuild is non-trivial, priced low enough not to provoke one) and the “minimum viable unit of saleable software” below which buying beats building. He’s betting his own job queue business on landing above that line. Required reading before you price anything.

Finally, Hanna Horvath’s I Make Good Money. Why Do I Still Feel Like This? zooms all the way out to the economy the rest of us are building software inside. It’s a sharp essay on the K-shaped economy and the “missing middle”: the American middle class as a deliberate postwar policy project that got unbundled and repriced piece by piece, splitting into “material precarity” (the basics slipping) and “positional precarity” (high earners whose expectations no longer resolve). The real line, she argues, isn’t income but capital (whether your money generates its own returns), and both groups misdirect their frustration sideways instead of up. US-centric, but the psychology of the treadmill travels. A good, humane palate cleanser after a lot of AI discourse.

End of batch.

The post Today’s recommendations: An agent is a model plus a harness, the coming loop, whether small software businesses survive AI, and why good money still feels like a treadmill appeared first on Cristian Nebunu.