Almost a Quarter of Skills Have No Numbered Steps
The workflow is the part of a skill that tells the agent what to do first. Nearly a quarter of the corpus does not have one, at least not in a form you can count.
Counting number-list items across 765 SKILL.md files:
| Number of numbered steps | Skills | Share |
|---|---|---|
| 0 | 183 | 23.9% |
| 1–5 | 148 | 19.3% |
| 6–10 | 332 | 43.4% |
| 11+ | 102 | 13.3% |
Median: 6 steps. Maximum: 37.
The correlation between step count and score is 0.17 — weak, and well inside our evaluator's noise. So this is not a "more steps, better skill" finding. It is a finding about the shape of the corpus.
The zero-step group is the interesting one
A skill with no numbered steps is not necessarily bad. Some are reference skills: they tell the agent how to think about a domain, or provide a compact protocol, and the "workflow" is implicit.
But many of the zero-step files read like documentation rather than instruction. They describe a topic — brand voice, positioning, a design system — without ever saying "do this, then this". The agent has to invent the sequence, which is exactly the work the skill was supposed to do.
The rubric scores this under workflow design (25 of 100 points). A file with no procedure cannot earn that dimension, regardless of how correct its prose is.
What six steps looks like
The median skill has six numbered steps, and the better ones share a shape:
- An input step that asks for missing information instead of guessing.
- Two to four transformation steps.
- A verification step — check the output before claiming done.
- An output step that states what the user sees.
- Zero: is the sequence obvious from the description? If not, add it.
- Six-ish: normal. Check that one of them verifies the result.
- More than fifteen: check whether the steps are at the right altitude. Very long step lists are often several workflows that should be separate skills.
Step 3 is the one that distinguishes a workflow from a recipe. It is also the one most often missing: only 42.9% of the corpus mentions verification at all.
The practical test
Open your SKILL.md and count the numbered items.
The number is not the goal. The goal is that an agent can execute the skill without inventing the order.