Vol. I · No. 01 ✦ Spring · MMXXVI
ISSUE · ONE · SPRING · TWO THOUSAND TWENTY-SIX
A Retrospective · In Ten Parts

Thirteen
years in the
Workshop

A field guide to the writing of Will Schenk & The Focus AI, from a Brooklyn Rails shop in twenty-fourteen to the agent infrastructure of a Connecticut lumber mill, two thousand and twenty-six.

§ · §
Will Schenk & thefocus.ai
340 posts · 13 years A Review in Ten Parts Printed on the Internet
I. Masthead

The workshop & the storefront.

Two sites, one author, three hundred and forty-odd posts, thirteen years.

willschenk.com is the workshop — a technologist's lab notebook running continuously since October of twenty-fourteen. Four post types (howto, labnotes, articles, and since twenty-twenty-three, fragments). It is where things are tried, broken, and figured out.

thefocus.ai/insights is the storefront — forty-two essays, use cases, and reports since November twenty-twenty-four, when Will announced "The Focus AI — I'm back." It is where things are named and framed for paying customers.

Same voice, two registers. This issue maps the full thirteen years and examines, in detail, what happened in twenty-twenty-four, twenty-twenty-five, and the first four months of twenty-twenty-six — the period in which a deployment-obsessed Ruby-and-Emacs craftsman became an agent-infrastructure theorist.

Contents

Ten Parts · Forty Pages
    01 The Thirteen-Year Panoramaa volume-by-year timeline p. 4 02 The Seven Erasdistinct technical phases p. 6 03 2024: Learning to Buildeighty-seven posts p. 12 04 2025: Consolidationfragments mature p. 18 05 2026: The Agentic Framephilosophizing > coding p. 22 06 The Focus AI Editorial Voicethree phases of a company blog p. 25 07 The Four Threadsthemes across the corpus p. 30 08 The Dual-Site Architecturehow the two sites relate p. 34 09 Field Notesartifacts, gaps, curiosities p. 37 10 Colophonsources & methods p. 40
Part One · The Thirteen-Year Panorama

Volume, year by year.

Each bar: the count of posts published on willschenk.com. Two spikes. One silence.

2014
16 posts
2015
10
2016
4
2017
8
2018
6
2019
15
2020
27
2021
33 · K8s peak
2022
2 · silence
2023
37 · return
2024
87 · the year
2025
44
2026
9 (ytd)
The Peaks Twenty-twenty-one — platform-engineer peak: Knative, faasd, OpenFaaS, Kubernetes + TLS, SQLite-in-browser. Twenty-twenty-four — the AI-stack speedrun, 2.6× any prior year.
The Silence Twenty-twenty-two has two posts, both published on January 2nd, then nothing until April of the following year. This is the pivot. Something ended.
The Return Twenty-twenty-three introduces a new form: the fragment. Short, reflective, opinionated. It is the vehicle by which public thinking re-enters the site.
II. Part Two · The Eras

Seven technical phases.

A site this long doesn't have one narrative. It has seven, each with its own dominant stack, aesthetic, and preoccupation.

I 2014 — 2015 26 posts

The Ruby-Rails Craftsman

Opens October 31, 2014 with How to track your coworkers — simple passive network surveillance. The voice is established immediately: terse, practical, occasionally mischievous. In the next six weeks alone: Bot Design Patterns, Scripting Twitter, Making a command line utility with gems and thor, Using rake for dataflow programming.

Stack: Ruby, Rails, Middleman, Bootstrap, Twitter API, Heroku. Articles are already reflective — Field of Dreams is 25 years old, Why Engineers build crappy products, Slow data and Fast Sites. The howtos teach; the articles editorialize.

II 2016 — 2017 12 posts

The Chromebook Philosophical Phase

Quiet years by volume, loud by content. Twenty-sixteen delivers the first properly philosophical piece: Yes, No, Maybe, Don't Know, WTF — the limits of Binary Logic and the Importance of Forgetting. The following year: Why Short Term memory is a thing: forgetting as a creative act — Philosophy meets computer science.

Meanwhile: three separate posts about living and coding on a Chromebook. React plus Material Design plus rmwc plus react-router as the working stack. A writer who is technical, yet already skeptical of complexity.

III 2018 6 posts

The JAMstack Interlude

All six posts are Hugo, Firebase, or serverless OAuth. The shortest era. Lays the groundwork for what comes next without yet declaring itself.

IV 2019 15 posts

The Decentralization Era

The year of IPFS. Setting up an IPFS Node, Setting up IPFS on a chromebook, Installing guix on IntelNUC, Setting up Indieweb Homepage, Controlling IKEA Tradfri devices, Reverse engineering APIs using Chrome Developer Tools, Terraform and Packer.

A clear posture: own the stack, own the data, skip the cloud where possible. This posture never leaves — it re-emerges, intact, when local LLMs arrive five years later.

V 2020 27 posts

COVID & Disposability

The biggest year so far. Leveraging disposability for exploration — how to play around without leaving a mess is the ur-text. Emacs Tramp tricks, Beginning Emacs (the rediscovery), Using Syncthing — who needs servers, IPFS and Fuse — the worlds data in your filesystem, Effigy, a distributed social data layer — Scuttlebutt is awesome, let's run with it.

The site now runs at a sustained cadence with a distinct aesthetic: lightweight, throwaway, own-your-data.

VI 2021 33 posts

The Kubernetes Peak

Peak platform-engineer. The stack: Knative, OpenFaaS, faasd, NextJS on Kubernetes with Prisma, Rails on Kubernetes with cert-manager, CloudEvents, docker-in-golang, Terraform everywhere. Also SQL in Org-Mode, Setting up emacs for typescript development, Wrapping a executable in a function, SQLite in the browser (as both howto and NextJS follow-up).

This is the final flowering of the build-all-the-plumbing-yourself worldview before the AI hiatus.

VII 2022 — 2023 39 posts

The Silence & the Return

Twenty-twenty-two has two posts. Both on January 2. Then nothing until April, twenty-twenty-three. The Rubicon year.

The return introduces a new format: fragments. First one: rivian trusts the driver (July 2, 2023). Thirteen fragments that year, on Rivian, AI panic, hallucinations, bad analogies, Plato on robots, threads-is-a-mess. A Taxonomy of AI Panic Facilitators (July 4) is the first properly argued AI post.

The writer who comes back is different. Still shipping (Kamal/Mrsk, Hetzner CLI, private docker registry, emacs blogging mode take 2) — but now also thinking in public.

§ · § · §
III. Part Three

2024: The Year of
Learning to Build.

Eighty-seven posts. Two-point-six times the next-biggest year. The site's defining year.

Forty-two labnotes, twenty-four howtos, twenty-one fragments, zero articles. (The labnotes-to-articles ratio tells the story: too busy shipping to write long.) Two parallel workstreams dominate: a ten-week speedrun through the modern LLM stack, and an eighteen-post-in-one-month push on EV charger infrastructure for Cornwall, Connecticut.

The entire modern LLM stack, speedrun from scratch, in ten weeks, published every step of the way. On The First Quarter of 2024

The Ollama discovery arc

It all turns on one post. February 27, 2024 — Running Google Gemma Locally: in which I discover ollama. That's the hinge of the whole modern period. Over the next twelve weeks:

Tab. I · The ten-week speedrun
Date Post Meaning
Feb 27 Running Google Gemma Locally Discovers ollama
Mar 29 Deploying ollama on fly.io Makes it remote
Mar 30 Programming with ollama Automates it
Apr 2 Geocoding with ollama Structured output
Apr 10 Streaming responses from ollama Real-time UX
Apr 16 Programmatically Interacting with LLMs RAG & function-calling reference
Apr 26 Basic LLM Chat UI A reusable UI
May 30 Chromadb on fly.io Vector DB infrastructure

The parallel Cornwall thread

Not everything was AI. While running the Ollama speedrun, Will was also shipping a real product — EV charger maps for Cornwall, Connecticut: Making a JSON api from a CSV file using fly, Deploying puppeteer on fly.io, Wait for the download to finish with puppeteer, Using Datasette to map out charger locations, Deploying OSMR (routing engine), POSSE rss to mastodon, Leaflet markers with vite build.

Eighteen posts in March 2024 alone. The productive peak of the entire thirteen-year archive.

The November 27th pivot

One line changes the frame: The Focus AI — I'm back (fragment). From this point forward the writing shifts. AI Guardrails, personal information edition (December 2nd). Unnecessary Knowledge — keep it lean (December 16th). Tone moves from "I figured out how to use an LLM" to "I'm shipping a product and thinking about its implications."

Feb 27, 2024 · Labnote
Running Google Gemma Locally
"in which I discover ollama" — the hinge of the whole modern period. Before this: a Rails-and-Kubernetes craftsman. After this: an LLM-native builder, shipping a post every few days.
IV. Part Four

2025: Consolidation.

Forty-four posts. Back to sustainable cadence. Nineteen of them fragments — the most ever. Long-form articles return for the first time since 2020.

Twenty-twenty-four was learning. Twenty-twenty-five is living inside the thing you learned. And writing more essays than tutorials for the first time in the site's history. Fourteen labnotes, eight howtos, nineteen fragments, three articles.

Feb 24, 2025 · Article
Living in mktemp and git and codespaces
"Everything is throw away until it isn't." The year's manifesto — the 2020 disposability thread updated for the AI era. Ephemeral workspaces are what let you move at LLM speed.

The keyboard-first project

A six-month quiet theme: rebuild computing around keyboard plus terminal plus AI, no mouse. Setting up OSX machine (Jan 25). Setting and using Aerospace (Jun 28). How to turn aerospace into an Application Launcher (Jun 29). Using Github CLI. Move to Mise — faster than asdf. Claude Code in Emacs — Integrating Anthropic's CLI with your editor (Aug 25). The last post is the fusion: two lifelong threads — Emacs and LLMs — explicitly merged.

The classical turn

Mid-year, the writing starts reaching backward in time. Emerson (Feb 27). The Solemn Silence of Written Words — Plato on Socrates (Jul 15) — the Phaedrus frame. Unreasonable Effectiveness of Compute — Moravec on the shortage of compute in 1976 (Dec 30). A writer looking for durable references for a fast-moving technology.

AI as subject, not tool

The fragments now function as a commentary track on the industry, unfolding in near-real-time: New o3 just dropped. Books in Gemini. Mistral AI is Le French. Knowledge Navigator — divergent futures. Talking to Cursor — great moments in AI. What we are most subtle in — thoughts on AI alignment. Agent Kickoff. Computer Held Accountable.

V. Part Five

2026 YTD:
The Agentic Frame.

Nine posts. One howto. Zero labnotes. One article. Seven fragments. The most essayistic the site has ever been.

The writer-to-tutorial ratio has inverted. The sole howto — Migrating to llama.cpp — drop ollama and run models directly, published April 16th — closes the arc opened two years prior. Ollama was the training wheels. Two years later, he's graduating past it, publicly.

The app-as-product era ends. Everything is agents operating on repos of data. After “Agents All the Way Down” · Feb 8, 2026
Tab. II · 2026 YTD · nine posts
Date Post Type Frame
Jan 7Jack Clark on TetragrammatonfragmentAnthropic on safety, deception
Feb 7Into the SingularityfragmentTime-horizon frame
Feb 8Agents all the way downarticleThe year's keystone thesis
Feb 13Lego-powered Submarine 4.0fragmentPhysical-world antidote
Feb 24Yelling at the ModelsfragmentUX frustration
Mar 5Shadows of GodfragmentTheological frame
Mar 6Lazzzored DodecahedronfragmentMaker-identity nostalgia
Mar 28The Will to Power IntelligencefragmentNietzschean frame
Apr 16Migrating to llama.cpphowtoCloses the 2024 arc

The title Agents all the way down is the single most compressed thesis statement of the current phase. By April, the writer has become a theorist who occasionally still builds.

VI. Part Six

The Focus AI
editorial voice.

Forty-two posts in eighteen months. Different register, same author. Twenty-three essays (fifty-five percent), six use cases, four model reports, and the beginnings of a paid practice.

Phase One

Everyday AI for Thinking

Nov 2024 — May 2025

Audience: curious professionals. Tone: demonstrative, "look what these tools can do."

Showpieces: Tools for thinking, Notebook LM for research, AI for research: DeepResearch a clear winner, How I classify models (first original taxonomy: small / foundation / educated), Clipboards are eating the world (a clipboard-POV allegory), The New Touch Interface, Agentic YOLO with Warp, Cursor, and Claude, Exposing Services with MCP (presciently early), Moral Vibe Check, Microsoft Build 2025.

Phase Two

Professional Practice

Jun — Sep 2025

Audience: developers deploying. Tone: opinionated, specific.

Showpieces: Geo-affordance, Feature Development on the go (build a PWA from your phone), June 2025 Coding Agent Report (the first benchmark-style report), Don't be passive aggressive with your agents, Technical Debt and the ROI Threshold (agents shrink future cost of messy code), GPT-5 and GPT-OSS, Code Generation with Local Models, Single file swift mini-apps (Swift as a scripting language).

Phase Three

Habitat OS & the Organizational Frame

Nov 2025 — Apr 2026

Audience: buyers. Tone: frameworked, benchmark-backed, philosophical.

Showpieces: AI Engineering Code Summit 2025, Weekend Coding Agent, AI tools fail loudly where humans failed quietly, "You're absolutely right" and other AI warning signs, How I Use AI in Jan 2026, Claude Code, not Code, The Data Flywheel Pattern, The Agent Habitat (keystone: agent = state + memory + git repo), The Car Wash Test (131 models, 76% fail), Sraffa's Gesture (Wittgenstein on LLMs), Can LLMs Use Real-World Tools? (Mercury-2, 39 models, Rivian data), Same Weights, Different Results, Gemma 4 on Your Machine, Building an Intelligent Organization, Laddering up from chatboxes to autonomy.

§ · § · §

The four thesis pillars

i

Agents, not chatbots

THE MATURITY CURVE

Laddering up, Building an Intelligent Organization, The Agent Habitat. Company name is Focus — focus = agent scope.

ii

Local models first

SOVEREIGNTY OVER THE STACK

GPT-OSS, Gemma 4, Mercury-2, Code Generation with Local Models, Image Gen on Apple Silicon. Tracks the llama.cpp migration on the personal site.

iii

Benchmarks as content

LEAD GEN BY SCIENCE

Car Wash Test, Mercury MCP, Gemma 4 Showdown, Same Weights Different Results. All powered by Umwelten — the benchmarks are the marketing.

iv

Philosophy as moat

HUMANITIES-SIDE DIFFERENTIATION

Sraffa/Wittgenstein, Moral Vibe Check, Stochastic Parrot. The humanities frame is load-bearing, not decorative. Mirrors the personal site's Emerson/Plato/Moravec turn.

VII. Part Seven

The four threads.

Four continuous themes, visible across the whole thirteen years.

I

The Craft Thread

2014 → PRESENT

Never stopped shipping. The stack rotates every two to three years: Middleman → Hugo → Astro. Heroku → Docker → K8s → fly.io → Hetzner. Rails → React → NextJS → buildless HTML + Server Actions. Bootstrap → Material → Tailwind + shadcn. asdf → mise. Ollama → llama.cpp.

What's remarkable isn't any one stack — it's that the motion never slows.

II

Disposability & Exploration

A SINGLE AESTHETIC, 13 YRS

Leveraging disposability for exploration (2020). Quick static site template — Rapid and disposable prototyping (2024). Living in mktemp and git and codespaces — Everything is throw away until it isn't (2025).

Tools: syncthing, terraform-templated servers, IPFS, Scuttlebutt, wormhole, codespaces, ephemeral git. Taste: lightweight, throwaway, own-your-data, no lock-in. This is why local LLMs clicked so hard in 2024.

III

Decentralization

A PERSISTENT POSTURE

2019-20: IPFS, indieweb, Scuttlebutt peak. 2023: private docker registry, gitolite, Hetzner CLI. 2025: ATProto article. Local models / llama.cpp are the current form.

The throughline: sovereignty over the computing stack. From IPFS to llama.cpp, same posture.

IV

The Philosophical Thread

DORMANT, THEN DOMINANT

Early signals: Why Engineers build crappy products (2015), Yes No Maybe (2016), Short Term memory (2017). Silent 2018-22. 2023 return as format: fragments invented July 2023. 2024 AI-philosophy cluster. 2025 classical turn (Plato, Emerson, Moravec). 2026 Nietzschean / theological / civilizational framing.

Paid-venue twin: Moral Vibe Check, Sraffa's Gesture. The thread went from private hobby to brand differentiator.

VIII. Part Eight

The dual-site architecture.

Two sites, one author, two jobs. The economic logic is exact.

willschenk.com
The Workshop
personal exploration · howto + labnote + fragment
"I figured out X"
│
The Focus AI hired, November 27, 2024
▼
thefocus.ai/insights
The Storefront
professional position · essay + use case + report
"Here's what X means for your organization"

willschenk.com never stopped — the personal cadence continued straight through the company launch. But from November 2024 onward, ideas that gelled there got upstreamed into formal essays on thefocus.ai.

The economic logic is clean: the benchmarks published on thefocus.ai are powered by Umwelten (the code in this very repo), which was built in public on willschenk.com. The personal blog is the R&D; the company blog is the market-facing layer; the repo is the load-bearing middle.

Tab. III · Register translation · seed → formalization
willschenk.com (the seed) thefocus.ai (the formalization)
Living in mktemp and git and codespaces (Feb 2025)Technical Debt and the ROI Threshold (Jul 2025)
Coding in one file (Jul 2024)Agentic YOLO (Mar 2025)
Thoughts on reading the llama 3.1 paper (Jul 2024)How I classify models (Jan 2025)
Claude Code in Emacs (Aug 2025)Claude Code, not Code (Jan 2026)
Agent Kickoff (Nov 2025)The Agent Habitat (Feb 2026)
IX. Part Nine

Field notes.

Artifacts, gaps, and curiosities worth recording.

The 2022 silence is real. Two posts all year: Using ActiveRecord outside of rails and Snowpack for fast builds, both on January 2nd. Then nothing until April 2023. Most-likely reading: a personal or professional reset between the Kubernetes era and the AI era.
Fragments are a post-2023 invention. All sixty fragments on the site postdate July 2023. They are the vehicle that lets the philosophical thread run hot — short enough to capture a half-thought, long enough to frame it.
The October 2025 gap on thefocus.ai is the only silent month on the company blog. Presumed Habitat OS build-crunch.
November 27, 2024 is the only date appearing on both sites simultaneously. The Focus AI — I'm back (fragment) on willschenk.com and Welcome to The Focus AI (essay) on thefocus.ai. The same day. The same announcement. Two registers.
The Astro migration is blogged from inside itself. Astro and Obsidian, Adding seo and social to astro, Separate git for blog writing all ship around September-November 2024. The publishing infrastructure is treated as a first-class artifact.
March 2024 is the single most productive month in the archive. Eighteen posts. Ollama experiments running in parallel with Cornwall charger infrastructure work. Peak parallel workstreams.
The word Umwelten never appears on willschenk.com. It is a Focus AI brand term and also the name of this repository. The split is intentional: the workshop doesn't name its tools; the storefront does.
There is no "year in review" post anywhere. The site never stops to summarize itself. That is — among other things — why this magazine exists.

Colophon

This issue was assembled from .firecrawl/willschenk-writing-index.md — the full 13-year post index from willschenk.com/writing, scraped via Firecrawl on April 21st, 2026 — together with .firecrawl/thefocus-ai-insights.md, the full 42-post index from thefocus.ai/insights, scraped the same day.

Topic clustering, year-over-year counts, and thread-tracing performed locally. Post counts rounded to the archive as scraped. "YTD" figures for 2026 cover January 1 through April 21. Any misattribution is the reviewer's alone.

Set in Italiana, Playfair Display, Cormorant Garamond, and JetBrains Mono — designed to feel printed, delivered via HTML.

★ Issue 01 · Spring 2026 · Finis ★