Methodology

How the library is built

Every pattern, example, and score on this site comes from a dated, sourced observation. This page documents the scoring rubric, the identification pipeline, and the bar a candidate behavior has to clear before it becomes a published pattern.

The Signal Score

Every captured item is scored 0–10 on four dimensions by an AI triage pass, then combined. Scores are triage machinery — they decide what reaches a human for approval and never appear on reader pages. This page is the only place they're discussed.

DimensionWhat it measures
RelevanceHow directly this bears on human–agent interaction design and agentic development.
NoveltyIs this a new behavior, or another instance of a known one? First sightings score high.
ShareabilityWould a designer or builder forward this to a colleague?
EvidenceScreenshots, shipped product, documentation — versus claims and demos.

≥ 7 story-worthy — a candidate for coverage · ≥ 8 captured into the library even without a story

The Identification Pipeline

  1. Sighting. Scouts (and readers) capture a dated, sourced observation. Summaries are written in our words; we never store source text beyond a short quote.
  2. Triage. Each sighting gets a one-sentence problem statement — "this solves the problem of ___" — and is tagged to existing patterns, or marked unclassified.
  3. Clustering. Unclassified sightings that describe the same problem accumulate. Three or more from at least two vendors become a named candidate on the Watching list.
  4. Promotion. Candidates that clear the promotion bar (below) get a full pattern page.
  5. Re-scoring. Published patterns are re-scored quarterly from sighting velocity: emerging → established → contested → fading. Status changes are logged in each pattern's changelog.

The Promotion Bar

A Watching candidate becomes a published pattern only when all of these are true:

  • ≥ 3 documented examples
  • ≥ 2 independent vendors
  • A one-sentence, product-agnostic problem statement
  • A screenshot or primary source per example
  • Distinct from every existing pattern (checked, not vibes)

Naming Rules

  • Name the behavior, not the technology
  • Two words or fewer
  • No vendor terms — vendor vocabulary is recorded per pattern under "what vendors call it"

Sources & Copyright

Sightings come from official documentation, changelogs, launches, demos, community discussion, and reader submissions. Every claim links its source and carries the date the behavior was observed — not the date we scraped it. Summaries and annotations are our own words. Screenshots are stored for commentary and criticism with source attribution.