Workflow Is the Weakest Dimension in the New Batch

The forty new skills are reasonably well structured. They are much less reliably executable.

Here are the six-dimension averages for the batch.

Workflow is the lowest average and the widest spread. Thirteen of the forty skills score below 3.0 on it.

The distinction between structure and workflow matters. Structure asks whether the document has a recognizable shape: frontmatter, headings, sections an agent can parse. Workflow asks whether those sections tell the agent what to do, in what order, with what inputs and outputs, and what to do when a step fails.

A skill can pass the first test and fail the second. Several of the new files have a clean ## Overview and ## Instructions section, but the instructions are a list of capabilities rather than a sequence. They tell the agent what the tool can do, not how to run a task.

The low end

The thirteen skills below 3.0 on workflow include asana-automation at 1.5, calendar-automation at 2.0, clickup-automation at 1.5, sheets-automation at 2.5, google-slides at 2.5, and the thin e-commerce how-to files: conversion-rate-optimization, ecommerce-ab-testing, shopify-seo, and five more.

Most of them share a shape: a list of API endpoints or marketing tactics, with no ordered procedure and no verification step. The agent is told what exists, not what to do first.

The high end

At the other end, jira scores 9.5 on workflow, k12-lesson-plan-creation 9.0, and k12-lesson-differentiation 9.4. These files encode a sequence: check the environment, route to a backend, perform the task, verify the result, and stop for approval when the action is destructive.

That is the difference between a reference document and a skill. The rubric rewards the second.

For the full definition, see our methodology. The trigger side of the same problem is covered in the next post.