Skill123 評分前 50 名中有 29 項 Agent 技能來自三位作者，文章分析其做法

We evaluate agent skills for a living — 600+ scored on six dimensions at Skill123. Sort by score and the author column collapses: eleven skills from one name, ten from another, eight from a third. Three people hold more than half the top slots.

Our directory holds 637 skills from hundreds of authors. The median author ships one or two skills. But the top 50? Three authors hold twenty-nine of them:

• calesthio — 11 skills (a complete programmatic-video pipeline)
• JimLiu — 10 skills (a 21-skill bilingual content toolkit)
• Jeffallan — 8 skills (a memory architecture)

This isn't nepotism — we score blind, one SKILL.md at a time against a fixed rubric. And it isn't resources — a skill file is markdown. So we read everything these three shipped, looking for the common thread.

1. They write systems, not files

Each author has a thesis about how agents should work, expressed across files. The video author shipped a router, a composition contract, animation rules, media handling, a CLI loop — a designed hierarchy where specialists assume the contract skill is loaded. The toolkit author shipped a full bilingual content pipeline: fetch → translate → diagram → slides → publish to WeChat. The memory author shipped capture, organize, recall, connect — one model of memory, four components.

Strip the branding and it's the same shape: a coherent system where each skill is a component.

2. Their floor is high, not just their ceiling

The video family runs 5.7–9.7 (one outlier). The toolkit runs 8.9–9.5. The memory set runs 9.0+.

The interesting number isn't the 9.7 — plenty of one-hit authors have those. It's the floor. A 9.7 with a 5.7 sibling is one great skill. A family that never drops below 8.9 is a process — the author has a way of writing skills that survives not caring about that particular one.

3. They win on the dimension everyone else loses

Across our whole directory, the weakest dimension is trigger quality — the "use when" clause that decides whether a skill ever fires (7.74 average). It's exactly where these three spend their effort. Their triggers don't describe the skill; they discriminate — each "use when" is specific enough to win the loading decision against the author's own other skills.

That's the counterintuitive part: shipping twenty skills forces the trigger discipline that makes all twenty good. Average authors write triggers like feature lists; these authors write them like tiebreakers. The portfolio shape drives the quality, not the other way around.

The recipe (no 90k stars required)

If you're writing skills:

1. Pick a domain you know deeply enough to know where one job ends and the next begins
2. Define the job boundaries before writing any file
3. Write each skill's trigger so it could not be satisfied by a sibling
4. Keep going until the family is complete — the discipline compounds

And if you're browsing: author is a quality signal, category isn't. Found one skill you like? Check the author's portfolio before searching the category. The odds of adjacent quality are dramatically higher — we measured it.

The full breakdown — including the anthropics counterexample and why the most official skills are the most modest — is at skill123.me/blog/super-authors. Every author's portfolio is linked from their skill pages.