The catalog just jumped from 155 to 206 patterns. This round pulled in 51 new ones from three repositories — and shipped a searchable, filterable pattern browser so all 206 are actually easy to get around.

🔍 Browse all 206 patterns → — search + filter by category / source / confidence, EN & 中文 in one toggle.

Where the 51 came from

Three source repos, mined by parallel research agents and deduped against the existing 155:

Source New patterns Flavor
superpowers 18 (156–173) Subagent-driven review orchestration + reviewer anti-gaming
claude-plugins-official 15 (174–188) Migration equivalence-checking + supply-chain security
mattpocock/skills 18 (189–206) Skill-authoring theory + frontier planning

Two of the four repos we scanned (awesome-claude-skills, andrej-karpathy-skills) hadn’t seen a single commit since the 2026-05 harvest, so they gave us nothing new. Honest accounting beats padding the count.

A new category: Skill Authoring

The mattpocock haul stood out enough to earn a new category — Skill Authoring — patterns about how to write skills, rather than what they do:

  • Leading Word (Leitwort) — anchor a whole region of behavior by reusing a compact pretrained concept (lesson, fog of war, tracer bullets) as a behavioral hook.
  • Invocation Economics — model-invoked pays in context load (the description sits in the window every turn); user-invoked pays in cognitive load (you have to remember it exists).
  • Sequence Cut — tuck post-completion steps behind a context boundary so the agent doesn’t rush the step right in front of it.
  • No-Op Test — go hunting for sentences the model already obeys by default, and delete them. Settle it by running the skill, not by arguing.

Highlights from the other two

  • Reviewer anti-gaming (superpowers) — a coherent family: a dispatcher must not pre-judge a finding’s severity or tell a reviewer what not to flag; an implementer’s self-justification (“left it per YAGNI”) never downgrades a finding; a plan-mandated defect is still a finding.
  • Migration equivalence (claude-plugins-official) — prove a version uplift changed nothing by running the same suite on both runtimes and diffing per-test, reproducing the baseline including its failures — not chasing “all green”.
  • Cross-service credential-routing detection — a credential read is malicious only when it’s routed to a different service than it belongs to; judge ownership by name, not by claimed use.

What’s next

The browser and metadata ship today. Full GOOD/BAD write-ups for the 51 new patterns are still in progress, so their detail links currently read write-up pending. The 155 existing patterns keep their full write-ups.

See the full 1–155 index or just browse everything.