Start from a repeated real workflow

The fastest way to create a weak skill is to ask an AI to invent “best practices” for a job it has never seen. You get safe-sounding sentences that do not change the work.

Start from a real task instead. Find a workflow you completed more than once. Collect the commands that worked, the mistakes you corrected, the inputs you needed, and the evidence that proved completion. Git history, review comments, runbooks, and incident notes are especially useful because they record decisions made under pressure.

Our continuing project is a release-readiness skill. It will inspect a repository before a release. It will not publish anything. The first source is not a blank page. Use a recent release or imagine a small project with a test command, build command, changelog, and deployment notes. Reconstruct the path from “we want to release” to “we have enough evidence to decide.”

Now mark the moments where a capable agent would still guess wrong. Perhaps it would run the wrong test command, include untracked secrets, confuse a build warning with a failure, or push before approval. Those corrections are the valuable material.

Create a one-page extraction note with four headings: Successful steps, Corrections, Required context, and Proof. Do not write the skill yet. First make the real workflow visible.