Why Developer Tools Score 9.21 and Documents Score 8.26 — A Category Gap Autopsy

We've said category averages are mostly about who wrote the code, not what the code is for. That's still true inside categories. But between categories there's a persistent, structural gap — and it has a cause you can act on.

Across our top-50 per category, the averages run:

We've made the point before that intra-category spread dwarfs these gaps — the best and worst skills within Documents are 4+ points apart. But a full point of average difference between Developer Tools and Documents is real, and it survives author-controlled comparisons. Here's what's actually driving it.

The stable-domain advantage

The highest-scoring categories wrap stable, well-documented domains with verifiable ground truth:

• Developer Tools wraps APIs, CLIs, and language ecosystems. The facts are documented; the failure modes are known; correctness is testable. A skill that says "use psutil for this, here's the exact call" can be right in a way a skill about "how to write good marketing copy" can never be.
• Productivity wraps thinking patterns — checklists, audits, verification gates. The domain is small and self-defined: the skill is the specification.
• Documents wraps file formats — also stable. So why does it score lowest?

The Documents anomaly

Documents has the highest ceiling in the directory (xlsx 9.8, docx/pptx 9.7 — the official anthropics trio) and the lowest average (8.26). The explanation: the top is written by the format's owner; the long tail is written by everyone else.

Below the official trio sit dozens of community document skills that share one weakness: they describe documents rather than tasks. "Word document processing" is a topic; "turn this markdown into a formatted monthly report" is a task. Our workflow dimension — the heaviest at 25 points — scores task-shaped skills, not topic-shaped ones. A category whose community skills are organized around a format rather than a job will systematically underperform on it.

The same shape shows up in Design (8.44): many skills describe capabilities ("generate images") rather than jobs ("make an article cover that fits our house style"). The exceptions score near the top — baoyu-cover-image (9.2) is exactly a job-shaped skill.

What this means if you're writing a skill

The gap is actionable in both directions:

If your domain is stable (APIs, formats, tools) — you have no excuse. The ground truth is written down somewhere; your job is to encode it precisely, with the failure modes and the version gotchas. Skills in these domains should be scoring 9+, and ours mostly do.

If your domain is mushy (copy, design, strategy) — the fix is to shrink the domain until it's checkable. Not "write a LinkedIn post" but "draft a post from these raw materials using one of 20 structures, with the hook in the first line." The highest-scoring writing skills all do this — they turn judgment into procedure, which is exactly what a rendering contract does for visual output.

The lesson generalizes: a skill's score ceiling is set by how verifiable its domain is — and the skill author's job is to raise verifiability by writing checkable contracts into mushy domains.

A prediction

As our new skill builder gets used, we expect the built skills to skew toward the stable-domain pattern — people build skills for their own repetitive, well-specified workflows first. The mushy-domain categories will keep importing their quality from the top authors. Both effects should widen the intra-category spread, which is exactly where the signal lives.