-
Seven decisions, one afternoon
A first agentic setup
The atlas has 1,294 projects in it, which is the wrong number to start from. This is one working setup instead: an agent, the things it loads, the place it runs, and the two tools that tell you afterwards whether it did what you asked…
-
Stated support, not inferred
Runs on Windows, and says so
Most agentic tooling is written on a Mac and tested on Linux, and "it's Python, it'll be fine" is how a Windows afternoon disappears. The atlas records a per-platform verdict for every project, and it distinguishes stated support from…
-
What to install after the agent
The Claude Code power kit
Claude Code is the harness the largest single share of this atlas is built for. Installing it is ten seconds; the next hour is the part nobody writes down. These seven cover the format skills come in, three bundles worth reading before…
-
A local stack, top to bottom
Nothing leaves the laptop
Sometimes the requirement is not cost or latency but that the code, the documents and the prompts stay on hardware you own. That rules out most of the atlas. What is left is a genuine stack: an engine, a runner in front of it, a chat…
-
Traces, tests and scans
Find out what the agent actually did
An agent that fails loudly is a good afternoon. The expensive failures are the quiet ones: the run that looked fine, the change that passed review, the skill that did something you did not read. None of this category is exciting and all…
How these are made
The picks live in
one JSON file, so a
disagreement is a diff rather than an argument about a rendered page. Every pick is checked against
the committed snapshot before anything is published: a project that has left the atlas, or a
"runs on Windows" pick whose platform verdict has dropped from stated to inferred, fails the build
instead of quietly weakening the page. Each set is also a link — the atlas reads
#list= in a URL, so a collection opens as a filtered view you can save,
print, or export as Markdown or HTML and take somewhere else.