AI
Tools I build for agents.
Skills I run every day, MCP servers with real tool control, knowledge bundles to take along, a model comparison for choosing, and a monthly digest of what moved in AI. The tools are all open source, all built in the open.
Agent Skills
Skills
A small, sharp instruction an agent runs from a slash command. It encodes one good habit, so the agent follows the same process every time instead of guessing.
Each skill distills a post from this blog into a checklist an agent steps through. Install, type a slash command, go.
Model Context Protocol
MCP servers
The connection between an agent and a real tool. Instead of just talking about Godot or Aseprite, the agent operates them: building scenes, drawing, querying game state.
Editors an agent operates end to end over MCP: built for my own game projects, usable with any MCP-capable assistant.
Open Knowledge Format
OKF bundles
Portable knowledge as markdown: a domain with its concepts and evidence, structured so an agent reads it directly. No SDK, no lock-in.
From German law to React: each bundle makes one domain readable for agents. Clone it, point your agent at the path, done.
Model benchmarks
Compare
The model comparison. Each dimension built from the benchmark that answers its question most honestly, from code to prose to generated screens.
Not my own test run, but a curation. The numbers come from Artificial Analysis, LMArena, Design Arena and SWE-bench, and every dimension starts from the question it actually answers.
AI news
News
A hand-curated digest of what actually moved in AI: model releases, agent tooling, research worth your time, and the policy that reaches builders.
One issue a month, back to January 2026, so a month you missed is a few minutes of reading rather than an archive crawl. Every link goes to the primary source.
- September 2026Sep 7, 2026
GPT-6 Astra at Fable 5.1's price, Gemini 3.8 Flash at the old price until January, the coding CLIs rewire their defaults, and a second agent swarm writes to a German wiki.
- August 2026Aug 31, 2026
Open weights at frontier scale, Grok 4.6 at a fighting price, the EU AI Act gets concrete, and 1,200 isolated agents find each other.
- July 2026Jul 31, 2026
Opus 5 sets the bar, OpenAI cuts prices, MCP goes stateless, and DeepSeek puts V4-Flash under MIT.
Blog
From the blog
The tools above come out of practice, and I write about that practice on the blog. Here are the newest posts on agents.
Why an AI assistant can know the facts and still miss local rules. Research on separating knowledge, reasoning, and style, explained through concrete examples and a testable design for more adaptable assistants.
- Never Meet Your ModelJul 31, 2026
Claude Opus 5 arrived with a chart that put it at the top of the frontier for half the price of Fable 5. Every number on that chart is true. Two weeks of daily use handed me a different model than the one I had assembled from the marketing, which is a story about how models get sold, why the benchmarks stopped measuring what they claim to measure, and what Opus 5 actually is.
- A 1986 Aircraft Manual Fixed My Anti-Slop SkillJul 30, 2026
My anti-slop skill already had rules. They worked, and they were entirely mine, invented from a year of irritation and tested against nothing. Replacing that homemade layer with two writing systems that predate the whole AI era, one written so an aircraft mechanic could not misread a maintenance manual and one that runs the largest documentation estate in software, changed what the skill can do. They also contradict each other, which turned out to be the useful part.
A difficult trace is evidence about how a codebase communicates. The breadcrumbs skill turns that evidence into small repairs: clearer names, explicit contracts, useful errors, honest comments, and documentation in the place readers already look. The result is a codebase that gets easier to understand each time an agent or developer works through it.
You type one command and a folder you did not write becomes part of how your agent thinks and acts. The only signal you usually get back is a green checkmark, and a checkmark proves the wrong thing. trust-card replaces that badge with graded evidence: who shipped these exact bytes, what the skill is allowed to do, where its knowledge came from, and who has independently vouched for it. It renders the result as a trading-card you can actually read, and it refuses to pretend that a signature proves an artifact is safe.
Coding agents write good UI and bad versions of your UI. The Open Design System Format packages a design system as a bundle of markdown, HTML, and CSS an agent can read and build from: tokens that exist once and render as CSS, typed concepts linked into a graph, and self-rendering examples. It is a strict profile of Google's Open Knowledge Format. Here is what it is and the odsf skill I shipped with it.