The book teaches the part that doesn't change. This page carries the part that does — the current documentation for every setting, screen, and price the book points at, chapter by chapter. Plus the department skills, free to download.
Each chapter of the book ends with a pointer to the matching section here.
Model and mode pickers for all four tools, and how each one wants to be prompted.
Running the five-minute audit: where the controls sit today, and what an admin may have switched off.
The exact settings paths — account-wide instructions, memory, and incognito — for all four tools.
Build screens for Projects, Notebooks, and Gems — the place a body of work lives.
Packaging a standard procedure so the tool runs it the same way every time. A working SKILL.md to copy is further down this page.
The fastest-changing chapter in the book. Agent availability moves month to month — check these before you rely on anything printed.
App-by-app coverage changes monthly and is impossible to print accurately. These are the live matrices.
Prices are deliberately not printed in the book, because they move. Each vendor's own plan page is the only source worth quoting.
Retention settings for all four tools. The one-page AI policy template is further down this page.
The same standard procedures the book teaches you to build, already packaged. They run on whatever you've got, including the free tiers. Each download contains the skill file, a one-page trigger card, and a Copilot prompt you can paste as-is.
Reconciles a bank statement against a GL cash account and returns an exception report. Never resolves anything on its own — every difference is flagged for a human.
DownloadTurns an aging export into a prioritised collections worklist with draft outreach. It never sends a message, adjusts a balance, or writes anything off.
DownloadMatches invoices against purchase orders and receiving records at the line level, and categorises every exception. It never approves or posts a payment.
DownloadTurns a bank or card transaction export into a categorised profit and loss with monthly and year-to-date views. It never guesses a category silently.
DownloadCopy this, change the procedure to one of your own, and save it as SKILL.md. Everything that makes it work is structural: the trigger wording, the numbered procedure, the line it will not cross, and the rule for missing data.
---
name: invoice-coding-check
description: Check coded supplier invoices against the chart of accounts and
prior-period coding, and return an exception list. Use whenever the user asks
to review, check, or sanity-check invoice coding, GL coding, expense coding,
or account assignment, or uploads a coded invoice register and asks whether
the coding looks right. Do NOT use for three-way matching or payment approval.
---
# Invoice coding check
## What this does
Compares how each invoice was coded against (a) the chart of accounts and
(b) how the same vendor was coded in prior periods, then reports only the
lines that disagree.
## What this never does
- Never re-codes an invoice.
- Never posts, approves, or releases anything.
- Never invents an account number that is not in the supplied chart.
## Before you run
Show the user these settings and let them change any of them in plain English:
- Look-back window for prior coding: 12 months (default)
- Confidence threshold for flagging: flag when the vendor was coded to a
different account in more than 20% of prior invoices
- Materiality floor: ignore lines under $100 (default)
Accept the user's files in whatever shape they arrive. Column names do not
matter — read the headers, map them, and confirm the mapping before running.
## Procedure
1. Read the invoice register and the chart of accounts.
2. Confirm the column mapping with the user before doing any work.
3. For each invoice line, find the account it was coded to.
4. Flag the line if ANY of these are true:
a. The account does not exist in the chart of accounts.
b. The account is marked inactive or "do not post."
c. The vendor was coded to a different account on the majority of prior
invoices inside the look-back window.
d. The line is coded to a suspense, clearing, or "ask accounting" account.
5. Do not flag lines below the materiality floor unless (a) or (b) applies.
6. Sort exceptions by dollar value, largest first.
## When data is missing
If the chart of accounts is missing, stop and ask for it. Do not infer a
chart from the invoices. If prior-period history is missing, run the check
with rules (a), (b) and (d) only, and say plainly in the report that the
history test was skipped.
## Output
A short report with:
- A one-line summary: lines checked, lines flagged, dollars flagged.
- A table of exceptions: invoice, vendor, amount, coded account, reason.
- A settings footer recording every default the user changed.
- A closing line: every exception is for human review; nothing was changed.
Most small companies do not need a twelve-page policy. They need one page that answers five questions, and a named person who owns it. Adapt this, take it to whoever signs things, and put a review date on it.
AI USE POLICY — [COMPANY NAME] Effective [DATE] · Owner: [NAME, TITLE] · Review date: [DATE + 6 MONTHS] 1. WHICH TOOLS ARE APPROVED Approved for company work: [list the specific tools and the specific accounts — e.g. "Microsoft 365 Copilot, signed in with your work account"]. Anything not on this list is not approved for company work, including personal accounts of approved tools. If you want something added, ask [NAME] rather than using it and hoping. 2. WHAT MAY NEVER GO IN Do not put any of the following into any AI tool: - Customer or employee personal information - Anything under a confidentiality agreement - Payment card, bank, or credentialing data - Anything covered by [INDUSTRY REGULATION, IF ANY] - Unreleased financial results, legal matters, or personnel decisions If you are unsure, the answer is no until you have asked [NAME]. 3. WHO CHECKS THE OUTPUT The person who used the tool is responsible for what it produced. AI output is a draft, not a result. Before anything leaves this company or enters a system of record, a named human verifies it against the source. No exceptions for speed. 4. WHERE AI MAY NOT DECIDE AI may draft, summarise, sort, and check. It may not make a final decision on: hiring or termination, credit or collections outcomes, pricing to a customer, safety or quality dispositions, or anything with a legal or regulatory consequence. A person decides and signs. 5. WHAT TO DO WHEN IT GOES WRONG If confidential information was entered into an unapproved tool, or wrong AI-generated information reached a customer or a record, tell [NAME] the same day. Reporting it promptly is the expected behaviour and will not be treated as a disciplinary matter on its own. Signed: ______________________ Date: ____________
If a link here is dead or a screen has changed, tell us and we'll update this page. That is the entire point of it.
Back to The Vinyl Academy