Agent Tutorials

Reading about AI is one thing. This page is the other thing: four walkthroughs, starting with the easiest possible way to try a skill — copy, paste, done.

Tutorial 1: Copy a skill into Claude or ChatGPT

The easiest way to try a skill. Nothing to install, nothing technical — if you can copy and paste, you can do this.

Step 1 — pick a skill. Go to the Skills Library and choose one that matches a task you did this month. First-timers: clean-credit-debit-amounts is the classic.

Step 2 — copy the whole skill page and paste it into a new chat in Claude or ChatGPT with this framing:

Here is a procedure document. Follow it exactly on the file I attach. Report the checks at the end, and list anything unusual instead of fixing it silently.

Step 3 — attach your file and let it work. Remove confidential information first — names, account numbers, anything you wouldn’t put in an email outside the office.

Step 4 — review the result. Do the totals still match? What ended up on the exceptions list? Tutorial 4 is the checklist.

Tip: in claude.ai you can save skills to a Project, or in ChatGPT to a custom GPT, so every new chat already knows your procedures. That’s a mini knowledge vault — which brings us to the real thing.

Then come tell the community how it went — or which task you wish had a skill.

Tutorial 2: Give a skill to an AI coding agent

The step up from Tutorial 1, for when you’re comfortable with slightly more technical tools. A coding agent works directly on your computer: it reads your files, does the work, and shows you everything it did.

You need: an AI coding agent (this walkthrough uses Claude Code; the pattern is the same in others) and one messy export you know well. Remove confidential information first.

Step 1 — pick the skill, same as Tutorial 1.

Step 2 — install it. Every skill page links to its raw text. Save it in your project like this:

my-close-work/
├── .claude/
│   └── skills/
│       └── clean-credit-debit-amounts/
│           └── SKILL.md      ← the file you downloaded
└── data/
    └── gl_export_march.xlsx  ← your sanitized file

That folder convention is how the agent discovers what it knows how to do. The skill’s description line is what it reads to decide when the skill applies — which is why every PANDAUDIT skill leads with one.

Step 3 — delegate, specifically. Start the agent in that folder and give it a memo, not a wish:

Use the clean-credit-debit-amounts skill on data/gl_export_march.xlsx. The amount column is “Net Amt”. Control total per the source system is $2,914,003.17. Give me the cleaned file, the exception list, and the before/after totals.

Step 4 — watch what it does. The agent will read the skill, do the work, and report back. Notice it follows the skill’s own checks — that’s the point. The procedure you’d give a junior is now something a machine can run.

Step 5 — review. Don’t skip to “looks good.” Use Tutorial 4.

Tutorial 3: Build your finance knowledge vault (the Obsidian way)

Obsidian popularized a simple, powerful idea: your knowledge lives in a vault — a plain folder of .md files that link to each other with [[wiki-links]]. No database, no lock-in, readable forever.

That same idea is exactly what an AI knowledge base is. A vault of markdown files is something both you and an AI agent can search, cite, and follow. Skills are just the executable subset of the vault — procedures instead of notes.

Step 1 — make the vault. A folder. That’s it. (Open it in Obsidian if you want the nice graph and linking; any editor works.)

finance-vault/
├── procedures/          ← skills live here
│   ├── clean-credit-debit-amounts.md
│   └── bank-rec-monthly.md
├── reference/
│   ├── chart-of-accounts-logic.md
│   ├── fiscal-calendar.md
│   └── vendor-aliases.md
├── decisions/
│   └── 2026-03-writeoff-threshold.md
└── close/
    └── close-checklist.md

Step 2 — seed it with three files. Don’t plan a taxonomy; write the three documents you already re-explain most often:

1. Your account-classification rules (the ones living in your head).
2. One recurring reconciliation, as numbered steps.
3. Your fiscal-calendar definitions (start with the fiscal-period-calculations skill).

Step 3 — link, don’t file. Inside a note, reference others with [[vendor-aliases]]-style links. When your bank-rec procedure says “normalize names per [[vendor-aliases]]”, both a new hire and an agent can follow the trail. Linking is what turns a pile of notes into a knowledge bank.

Step 4 — point your agent at it. With a coding agent, open it in the vault folder — it can read every note. With a chat AI, add the vault files to a Project / custom GPT. Then ask something you already know the answer to and check that it cites the right note. No citation, no reliance.

Step 5 — grow it by subtraction. Every time you answer a question twice, the second answer becomes a vault note. Every task you do twice becomes a skill in procedures/. A year from now the vault, not any one workbook, is your most valuable file.

Tutorial 4: Review agent output like a workpaper

The agent is a very fast preparer. You’re the reviewer. Same discipline as ever:

1. Tie the control totals first. Rows in = matched + exceptions. Amount in = amount out ± documented adjustments. If totals don’t tie, stop — nothing else matters.
2. Read the exception list, not the clean list. The agent’s job was to shrink your review to the rows needing judgment. Judge them.
3. Spot-check the “clean” side. Sample a few matched/cleaned rows back to source. Trust builds from verification, not vibes.
4. Check the skill was actually followed. The skill’s Validation section is your review checklist — the agent should have reported each item.
5. Keep the trail. Save the skill version, the agent’s output, and your notes together. That bundle is the workpaper.

Something break? That’s genuinely useful — bring it to the Discord #agents channel and the fix becomes an update to the skill for everyone.

Where next

• Skills Library — pick your first skill
• AI Field Kit — the concepts behind all of this
• Discord — share results in #skills and #agents