Only 22.7% of Skills Say What the Output Looks Like

A workflow can be correct and still produce the wrong artifact. The output section is what prevents that.

Counts across 765 skills:

The remaining three quarters describe a process and leave the artifact implicit.

What happens without an output contract

The agent executes the workflow and stops at the last step. Whatever that step produced becomes the deliverable. If the last step is "refine the copy", the output is a paragraph. If it is "save the file", the output is a path — and the user never sees the content.

In practice, the agent's default is to summarize what it did. That summary is not the artifact the user wanted.

An output section makes three things explicit:

1. The artifact type — a table, a file, a message, a list.
2. Its structure — the sections or columns, in order.
3. The report-back — what the agent says after producing it.

A minimal template

## Output

Produce a Markdown table with one row per candidate:

| Column | Content |
|---|---|
| Name | Full name |
| Signal | One sentence, max 20 words |
| Confidence | high / medium / low |

After the table, state the number of rows and any candidates you excluded, with the reason.

That is twenty seconds of writing and removes most ambiguity about "done".

Output is not verification

These are adjacent and often confused:

• Output defines what the result is.
• Verification defines how to check that the result is correct.

A skill can have a perfect output template and no verification (the artifact is well-shaped but wrong), or verification with no template (the check passes but the user gets a summary). The best skills have both, as separate sections.

Why the number is low

Writing an output template requires deciding what the skill is for. A workflow can be assembled from steps; an output requires committing to a shape. That is harder, and it is also the part users judge.

If you are editing a skill, start at the end: write the artifact you want to receive, then make the workflow produce it. The 22.7% who did this report fewer "that's not what I meant" iterations — we cannot measure that directly, but the mechanism is not subtle.