The patterns we run on, packaged.
Our agents work from a library of skills — small, versioned reference documents that codify how a task is done well. These are the ones general enough to share. Each follows the open Agent Skills format: a SKILL.md any modern agent runtime can load, plus reference scripts where useful. MIT-licensed. Adapt freely.
A skill earns its place here the same way it earns its place in our library: it prevented a real failure, more than once, in production.
Deploys that refuse to break the site.
Replaces "build && rsync --delete" with four fail-closed gates: build, verify output, scrub forbidden strings, atomic swap. The live directory is only ever touched on the success path. Includes the full reference bash implementation.
RSS 2.0 + JSON Feed 1.1 + llms.txt, one substrate.
The three-format syndication pattern this site runs: classic RSS for the installed base, JSON Feed for modern aggregators and AI summarizers, llms.txt for language-model crawlers. All three reading from a single canonical data module so metadata never drifts.
LLM pipelines that survive malformed JSON.
LLMs fail structured output in two predictable ways: truncating mid-string at token limits, and returning a single string where you asked for an array. Two small Python helpers — type coercion and truncated-array salvage — that keep unattended pipelines from silently crashing.
One file, any runtime.
Unpack the archive into your agent's skills directory. The SKILL.md frontmatter carries the name and a one-line description your runtime uses to decide when to load it; the body carries the working knowledge. The format is supported by Claude Code, GitHub Copilot, Cursor, and most agent runtimes that adopted the Agent Skills standard.
No runtime? The files are plain Markdown — they read just as well as a checklist for a human operator.
Want a skills library like this for your operation?
This is a working sample of how we codify operating knowledge so agents and people share the same playbook.