Agent skill · Writing

SOP to skill

Turns a written process, checklist or recording transcript into a spec-valid SKILL.md: asks about missing decisions, writes the steps, rules, inputs and an example, and keeps the description to what it does and when to use it. Use when asked to turn a process, an SOP, a runbook or 'how I do X' into a skill.

The sop-to-skill/ folder with SKILL.md.

The skill itself. Agents read the name and description first and load the rest only when a task matches.

sop-to-skill/SKILL.md
---
name: sop-to-skill
description: "Turns a written process, checklist or recording transcript into a spec-valid SKILL.md: asks about missing decisions, writes the steps, rules, inputs and an example, and keeps the description to what it does and when to use it. Use when asked to turn a process, an SOP, a runbook or 'how I do X' into a skill."
---

# SOP to skill

Turn someone's process into a skill another agent can follow without asking them.

## Steps

1. Read the process. Write down, in order: the trigger (what starts it), the inputs it needs, each step, every decision point with its rule, and what "done" looks like.
2. List what's missing or ambiguous: decisions without a rule, steps that depend on tacit knowledge, tools or access the agent won't have. Ask about these, at most five questions in one message, each with your suggested answer.
3. Pick a name: lowercase letters, digits and single hyphens, at most 64 characters, describing the job ("monthly-close-checklist", not "finance-helper").
4. Write the description in one or two sentences: what the skill does, then "Use when …" with the requests that should load it. Keep it under 1,024 characters and specific enough that it won't load for unrelated work.
5. Write the body: a one-line purpose, Inputs, Steps (numbered, one action each), Rules (the decisions and limits), and Done when. Add one short example of the output.
6. Move long reference material (templates, policy text, lookup tables) into references/ files and link them from the body.
7. Check the result against the spec: frontmatter has only name and description, the name matches the folder, every referenced file exists, no step needs a tool the user didn't mention.

## Rules

- Keep the person's process. Don't add steps they don't do; list suggestions separately at the end.
- Write steps as instructions to the agent ("Check the invoice total against the PO"), not as descriptions ("The invoice is checked").
- No scripts, credentials or URLs to fetch unless the process already uses them, and then say so in Rules.

## Example

Input: "Every Friday I export open support tickets, tag the ones older than 7 days, and post the count per agent in #support-leads."
Output: a `weekly-stale-ticket-report` skill with inputs (the export), five steps, a rule for what counts as stale, and an example post.