Why SOPs Fail as Prompts

The problem with pasting SOPs into chat

Your team already wrote the procedure. It lives in a doc: numbered steps, an owner, maybe a tools list. So the obvious move is to paste it into an AI chat and ask it to "follow the SOP."

That works — once. Then it doesn't. The model skips step 4 on Tuesday, invents a new output format on Wednesday, and quietly uses the most expensive reasoning it has for a task that needed none of it. Nothing was wrong with the SOP. It was written for a human who fills in gaps with judgment. A model fills gaps with guesses.

What a machine actually needs

A machine-readable capability declares things a human SOP leaves implicit:

• Inputs and outputs — what comes in, what must come out, in what format.
• Steps as structure — numbered, granular, no "use your judgment" steps.
• Routing — which model tier each step actually needs.
• Cost estimates — expected tokens and runs per month, so spend is predictable before execution.

That's what SKILL.md is: an open spec that turns an SOP into a portable capability. Portable means the same file runs in any compliant runtime — Claude, OpenAI, Gemini — without a rewrite.

The pre-check below

The "Do It" panel runs the same lightweight readability pre-check the Cromus demo uses. It looks at structure: does the SOP have granular steps, a purpose statement, ownership, a tools list? It is not a cost or governance score — it only measures whether your SOP is structurally ready to compile. For the full structural validation, use the SKILL.md validator.

Paste one of your own SOPs, or pick one of the five industry scenarios.