Why
Both of these entered the queue for the same reason, and it is not the reason a reading list usually gets added. This report itself runs on a SKILL.md, so the value of Google's ~90 distilled-knowledge skills — product, architecture, best practices, packaged the same way Claude's SKILL.md ecosystem packages them — is not the content. It is comparing how Google structures skill format and trigger design against Claude's approach. A skill is a small design artefact with a hard constraint (it must be worth loading), and reading someone else's is the cheapest way to see the constraint solved differently.
The course is the same instruction at a different scale. A free ex-Google-engineer course spanning agent design, human handoff, RAG and vector DBs, Cloud deployment, WhatsApp monetization, conversation-to-lead conversion, and loops versus graphs. The value is the end-to-end arc — Agents → RAG → Deployment → Leads → Revenue in one continuous piece — sitting at the intersection of the LLM track and the sell agent services business category. Most material covers one of those stages well; almost none shows the whole shape.
And the honest note is that neither deserves more than thirty minutes, which is why they belong in one card rather than two. one-page-not-the-curriculum argues that graded material is a reference to open at today's block rather than a sequence to walk — and a reading list is exactly the thing that quietly assumes otherwise. Two entries, two half-hours, one instruction: take the shape and leave the rest.
How it works
What to take from each, and what to skip
| Google Skills Repository (~90 skills) | Free 3-hour agent course | |
|---|---|---|
| What it is | Distilled knowledge packaged as agent skills — product, architecture, best practice | Agent design → human handoff → RAG/vector DBs → Cloud deployment → WhatsApp monetization → leads → loops vs graphs |
| Why it is here | This report runs on a SKILL.md — format comparison, not content |
The end-to-end arc in one piece |
| Take | Trigger design, how much context a skill assumes, where it stops | The design intro and loops versus graphs |
| Skip | The other ~87 | Everything the two chapters already covered |
| Budget | 30 minutes | 30 minutes |
The one thing worth writing down afterwards
How does each decide when to fire? A skill's trigger is the only part that cannot be borrowed carelessly — it encodes an assumption about what the reader is doing at the moment the skill becomes relevant. Google's answer and Claude's answer to that question are the comparison, and one paragraph noting where they differ is the entire deliverable of this card.
Why loops versus graphs is the chapter to keep
It is the one structural decision in agent design that survives changing every other choice: a loop re-enters the same step until a condition holds; a graph moves through named states with explicit edges. The first is easier to write and harder to reason about; the second is the opposite. Everything about observability, retry behaviour and cost ceilings follows from it, which is why the-harness-not-the-model cares about it and a monetization chapter does not.
Where it lands in Jayverse
- Rabbit: decide loop vs graph for the payment agent explicitly. Given irreversible on-chain actions, model mandate execution as a graph of named states (proposed → simulated → approved → sent) rather than an open loop, for auditability.
- OFA: model solver bidding rounds as a graph, not an open loop. A named-state design keeps a failed bid from silently retrying into a duplicate, the same failure mode the loop-vs-graph chapter warns about.
- Auditor/gitboard: borrow the trigger-design discipline for internal runbooks. Any future SKILL.md-style runbook (e.g., Verex resolution) should state an explicit "when does this fire" line, the way a good skill trigger does.
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| steal the structure | 형식(구조)만 따다 쓰다(내용이 아니라) · 남의 자료에서 배울 부분을 짚을 때. "steal the structure, not the content" |
| distilled | 정제된·핵심만 추린 · 방대한 지식을 요약해 담아낸 것. "Ninety-odd distilled agent skills from Google" |
| sitting at the intersection of | ~의 교차점에 있는 · 두 분야가 겹치는 지점에 위치할 때. "sitting at the intersection of the LLM track" |
| spanning | ~을 아우르는·걸쳐 있는 · 강의나 자료가 여러 주제를 포괄할 때. "A free ex-Google-engineer course spanning agent design" |
| end-to-end arc | 처음부터 끝까지 이어지는 흐름 · 단편이 아니라 전체 과정을 보여줄 때. "The value is the end-to-end arc" |
| deserve more than | ~할 가치가 그 이상은 아니다(부정문에서) · 투입할 시간의 상한을 정할 때. "neither deserves more than thirty minutes" |
| quietly assumes otherwise | 은근슬쩍 다르게 전제하다 · 겉으론 안 그런 척하지만 실은 다른 가정을 깔고 있을 때. "a reading list is exactly the thing that quietly assumes" |
| cannot be borrowed carelessly | 함부로 빌려 쓸 수 없다 · 맥락 없이 그대로 가져다 쓰면 안 되는 것. "the only part that cannot be borrowed carelessly" |
| encode an assumption | 전제를 담고 있다(내포하다) · 설계 하나에 암묵적 가정이 녹아있을 때. "it encodes an assumption about what the reader is doing" |
| survive changing | (다른 걸 다 바꿔도) 그대로 남다·유효하다 · 다른 선택이 바뀌어도 변치 않는 핵심 결정. "survives changing every other choice" |