Library Skills, How To Actually Use Them

A skill is a reusable instruction set that makes the AI run a workflow the same reliable way every time, instead of you re-explaining it each session. Two ship here:

tax-research, turns Module 2 into a tool that retrieves and verifies primary sources for every citation and refuses to present authority it couldn't confirm.
vendor-vetting, vet an AI tool/vendor for putting client data into it. It researches the vendor's actual published terms (privacy policy, DPA, trust center, sub-processor list, SOC 2, data-residency), maps them to §7216 / FTC Safeguards / AICPA / Circular 230, and returns a decisive Approve / Conditional / Reject verdict tied to a specific use, defaulting to "No for client data" (with the exact gap to close) rather than sitting on the fence. Run it as /vendor-vetting [tool name], for [use], on [tier]. Pairs with the ai-use-checker (task-level §7216 call) and the vendor due-diligence checklist.

You don't need to be a developer. There are three ways to use either skill, depending on which AI you run.

Option A, Claude Code users (the real skill) ⭐ most capable

The skill is a folder with a SKILL.md. Claude Code loads it automatically.

Install (one time): copy the tax-research/ folder into either

your project's .claude/skills/, or
your user skills folder ~/.claude/skills/ (available in every project).

cp -R tax-research ~/.claude/skills/

Use it: in Claude Code, type

/tax-research Can an S-corp shareholder deduct home-office expenses for a separate trade or business?

Claude runs the full workflow, frames the issue, maps candidate authorities, uses web search to pull and quote the real Code/reg/ruling text, marks each cite ✅ verified or ⚠️ unverified, runs a devil's-advocate pass, and outputs a research memo. Because verification is built in, it won't hand you a confident fake citation.

(This skill is already installed in Charlie's ~/.claude/skills/tax-research/, just type /tax-research to run it.)

Option B, Claude.ai (web/desktop) users, a Project ⭐ best for non-coders

Claude.ai "Projects" let you save reusable instructions + reference files.

Create a new Project (e.g., "Tax Research, verified").
Open Project instructions / custom instructions and paste the body of SKILL.md (the part below the --- frontmatter, the rules, source stack, and procedure).
Optionally add 02-tax-research.md, REGULATORY-FOUNDATION.md, and the memo template as Project knowledge files.
Start any chat in that Project and ask your tax question. The instructions make every chat follow the verify-then-rely workflow. (Web search must be enabled for live verification.)

Option C, Any assistant (ChatGPT, Gemini, Copilot), a saved prompt

Paste the body of SKILL.md at the start of a chat (or save it as a ChatGPT "Custom GPT" instruction / saved prompt), then ask your question. You get the same workflow and guardrails; the quality of live verification depends on whether that tool can browse/retrieve.

What "using it" looks like, a quick example

You: /tax-research Is a SaaS subscription sold to a customer in Texas subject to Texas sales tax?

The skill: restates the issue → identifies Texas Tax Code + Comptroller rules as leads → fetches the actual statute/rule text and quotes it → flags anything it can't verify → gives a reasoned answer with a contrary view → outputs a memo, reminding you to verify and sign off.

The difference from raw chat: you never get a citation the tool didn't open and read first.

Important: this assists research, it does not replace you

Every skill here keeps the human as the reviewer of record. The output is research for your review, verify the authorities, apply judgment, and sign. And keep client identifiers/SSNs/EINs out of prompts unless you're on a firm-approved tool (see Guardrails and Regulatory Foundation).