How it works
A confidently wrong legal answer is worse than no tool at all. Everything about how these skills are made, and how you use them, follows from that one fact.
The two layers
A plain language model can sound authoritative and cite sources, which makes a wrong fact more damaging, not less. So every skill keeps two things separate and holds each to its own standard. The judgment is drafted by AI and reviewed by an expert, because the model is good at knowing how to approach a task and review catches the gaps. The authority comes from real law and is verified by an expert, because that is where a made-up statute would hide.
How to approach the task, the way an election expert would.
Expert-reviewed
The actual law, the facts the judgment grounds itself in.
Sourced from real law · expert-verified
Where the value comes from
A vanilla model can already write in an expert voice. What it cannot do on its own is be right about your law. Two things close that gap.
The skill makes the model work through the task the way an experienced official would: identify the governing rule, address every part of the request, write for a real reader, and stop to flag what is uncertain. This part works today.
The skill carries the actual law, sourced and dated, so the answer rests on something real instead of the model's memory. Instructions alone cannot do this. It takes verified references, and an expert to check them.
See the difference
Ask a plain model the same kind of question twice and you get two slightly different answers, with no way to know which one dropped a required element or misremembered a deadline. A skill removes that variance: it runs the same expert-checked discipline every time and binds the facts that matter to cited, dated law.
“Make this ballot-return instruction readable for voters.” Now imagine the same kind of ask a hundred times over, across notices and staff.
Sign your ballot certificate, seal it inside both envelopes, and mail it back by Election Day. Your signature is required for your vote to count.
What you can’t rely on
Return your ballot so the county board receives it by 8:00 p.m. on Election Day. Ballots that arrive after that are not counted (unless the law says otherwise). Then: sign the inner envelope, seal your ballot inside it, and place it in the outer envelope.
What holds, every time
Run it once or a thousand times, by anyone on staff: the skill applies the same verified discipline and grounds the facts in cited law. That repeatability is the point, not the keystrokes it saves.
Checking the work
For certain tasks, a model is a better reviewer than it is a writer. Those skills draft a response, test it against a short list of explicit, expert approved criteria, revise, and only show you what passes. A records response is checked for completeness, correct classification, and grounding in cited law. A plain language rewrite is checked for preserved meaning, retained requirements, and reading level. The loop is only added where a real test exists, because agreeing with yourself is not the same as being right.
How a skill is made
Each skill is drafted, sourced from real law, reviewed by a subject matter expert, tested in a real tool against real questions, then dated and published. The same lines hold every time:
Putting it to work
A skill is portable expertise, not an app. It goes wherever you already work. Pick whichever fits how you use AI today.
Copy the skill’s contents into Claude, ChatGPT, Copilot, or Gemini, then ask your question. Works everywhere, no setup.
Set it up as a Claude skill, a custom GPT, a Copilot agent, or a Gemini Gem so it is ready whenever you need it.
Grab the whole folder (SKILL.md, references, checks, examples), versioned and dated. Yours to keep and review.
Step by step
The download is a Claude skill: a SKILL.md file plus its reference material. Claude takes it as is. For the other tools you paste the contents of SKILL.md as the instructions and add the reference files as knowledge. Here is the path in each.
.zip. Claude reads the SKILL.md and shows the name and description.Skills on claude.ai are per person, not shared across an organization. In Claude Code, drop the unzipped folder into ~/.claude/skills/ instead of uploading.
chatgpt.com/create.SKILL.md into Instructions. There is an 8,000-character limit, so if it runs long, move the rest into a knowledge file.Building a custom GPT requires a paid ChatGPT plan.
microsoft365.com/chat or in Teams. You need a Microsoft 365 Copilot license.SKILL.md into Instructions.Copilot is the one tool that will not take loose file uploads as knowledge. The reference files have to live in SharePoint or OneDrive.
gemini.google.com and sign in.SKILL.md into the instructions.For a large set of reference files, Google suggests building a NotebookLM notebook and connecting it to the Gem.
Three habits
A grounded skill is only correct for the place whose law it cites. Tell it your state and county. When a skill is marked as working in any jurisdiction, it brings the method and leaves the local facts to you.
These skills are built to surface what they are unsure about: a date to confirm, an exemption to verify, a question for a lawyer. When a draft raises a flag, that is the point. Resolve it before you act.
Every result is a draft for your review and signature. It does not send anything and it does not replace your judgment. You remain the authority on your office and your jurisdiction.