Skills are the new style guide
A style guide is a document people are supposed to read. A skill is a set of instructions an agent actually follows, which makes packaging your team's judgment a different kind of work.
Nearly every design team has a style guide that nobody opens at the moment they need it. It lives in a wiki, it was accurate on the day it was written, and the people who most need to follow it are the least likely to look. A skill is a packaged set of instructions that an agent loads when a task calls for it, so the guidance arrives at the moment of the work and nobody has to remember to go find it. Intuit's content design team documents a similar setup in its public guidance on using AI, including a warning that the Markdown files used to ground its tools can fall out of date.
A skill built around a checker
I wrote a skill for the research step of a daily audio show on this site. It takes a short phrase, researches it, and prints a cited outline that a script-writing pipeline consumes. The skill opens with a section explaining why the outline is shaped the way it is, and that section carries most of its value. The downstream writer treats the pasted outline as its only source of facts, and a fidelity checker holds every sentence of the script to it, so each rule in the skill traces to a way that machine reader fails.
A quotation can be spoken on air only if the outline pastes the exact words inside double quotation marks, so every quote has to be copied word for word from the page it came from, and a reference without its text gets named and explained and never quoted. No quoted text may contain digits, because the script speaks numbers as words and a quotation with a year in it stops matching the outline and fails the check. Every numbered point ends with a bracketed citation, and each publisher cited needs a matching bullet in the sources list, spelled identically. The title is read aloud inside a sentence and also becomes the public title, so the skill specifies sentence case and how a leading article is handled. Each rule carries its reason in the skill's opening section, so an agent that meets a case the rules never covered can reason from the checker's behavior.
A verifier that runs the build's own checks
The skill ships with a script, verify-quotes.mjs, that checks an outline before it goes anywhere. It confirms that every quoted string appears on one of the cited source pages (compared the way the build compares, ignoring case, punctuation and curly quotes), that no quote contains a digit, and that citations and sources agree in both directions. It imports the build's own outline checks, so the verifier and the pipeline run the same code and cannot drift apart, and it exits 1 on any failure. A flag skips fetching the pages and checks format alone for a fast pass. The person pasting the outline then reads something a machine has already passed, and the skill's instructions and the build's behavior sit in one repository where a change to either shows up in the same pull request.
State a skill should not invent
The skill also holds a registry of numbers that are permanent. It tells the agent to read the registry before researching, reuse an existing number when the topic is already listed, take the next unused number for a new one, and ask when it is unclear whether a phrase is new or part of an existing entry. An agent given a counter and no rule for it will assign numbers confidently and wrongly, so the skill writes the rule down and names the one case where it should stop and ask.
What carries over to a design team
The same anatomy suits a design problem, and the following is advice for building one, written as a plan. Open the skill with a short section on why the team solves the problem the way it does, list the rules with a reason beside each, show two or three real examples good enough to copy, and ship a script that flags what a machine can see. Rules that can be checked get a checker, and rules that cannot get examples, which matches what Intuit recommends for prompts: clear examples, background context, key terms highlighted, and a note of what to avoid. Start from a real failure, since skills written from imagined problems come out generic. Give each skill a single job and an owner, and keep it in the repository next to the thing it describes, because instructions stored away from the pipeline drift and the agent follows the outdated version faithfully. Intuit grades model outputs against a shared human rubric and feeds the patterns back into its prompts, and that loop is worth copying for any skill that matters.
Be the first to rate this article.
Let's work together.
Open to select projects and collaborations — design systems, accessibility, and AI-native product work.