Skills vs Agents: Where Judgment Belongs in an AI Workflow

Skills vs agents, in plain terms: the agent is the AI that does the work and makes the calls. A skill is the written procedure it follows. A script is a small program inside the skill that runs a check the same way every time. Good AI work keeps moving steps from the agent into the skill, then into scripts.

If your team already uses ChatGPT, Claude or Copilot, you are probably at the first layer: a prompt, an answer, and a person checking the result. That's a fine start, but the AI redoes the same steps on every run and can get any of them wrong.

Picture a restaurant kitchen. The agent is the cook. The skill is the recipe for a single dish, which the cook reads when that dish is ordered. The script is the kitchen scale. It doesn't interpret anything: the recipe says 200 grams of flour, and the scale makes sure it is 200 grams.

In tools like Claude, GitHub Copilot and Codex, a skill is a folder with written instructions and often a few small programs. Both travel in the same package, but they work in very different ways. The AI reads the instructions and interprets them, so it can still misread a step. The computer runs the scripts, so they give the same answer every time. Engineers call that deterministic: same input, same result.

Ask of each step: could I write down the check? If you could, the AI should not be doing that step from memory.

1. If the step has one right answer, it becomes a script. Do two totals match? Is this date in the right month? Is this word on our list of words we don't use?
2. If it is a known procedure with a few decisions along the way, it goes in the skill's instructions: where the documents are, what order to do things in, how to record the result.
3. If it means weighing options with incomplete information, it stays with the agent. Is this unusual number a mistake or a real change? This is the only kind of step the agent should keep.

Our bookkeeping agent attaches the receipt or invoice to each transaction in our books and checks that the sales taxes are split correctly. Every business has this kind of grunt work. In one August run, it took our list of transactions missing a receipt from 18 to zero.

1. A saved search query finds every transaction with no receipt. The agent runs it at the start and again at the end, so "done" is counted, not remembered.
2. The amount on the receipt must match the transaction.
3. The GST and QST on the receipt must match the tax lines in the books, and the entry must still balance after the split.
4. An email with three receipts must touch three transactions. Before this check existed, one run attached two out of three.

Each vendor sends receipts differently: a link in the email, an attachment, or the merchant's own system. The agent attaches the real document, never a link that will stop working in a few months. If it can't find the file, it leaves the transaction empty and says so. Some vendors send one document and no receipt after that, like a financing contract that authorizes a pre-authorized payment every month. There, the bank statement, backed by that contract, is the proof, and the agent notes it so the next run doesn't go looking for a monthly receipt that will never come.

Which account does an expense belong to? Are two identical charges a duplicate or two real purchases? The agent proposes an answer with its reasons and flags it. A person approves or changes it.

The instructions once said "flag, never correct," with no limit on what that covered. So when the agent found two invoices booked without their Quebec sales tax line, it reported them and stopped. Looking wider, we found six transactions in the same state: $39.92 of QST we could have claimed back, unclaimed for up to four months. Transactions created from a card statement have no tax lines, which is correct at that point, but nothing told the agent to split them again once the receipt arrived. June got fixed because someone noticed, but no rule came out of it. In July, the error came back: the fix had been applied to the past transactions, and no check was added to catch the next ones.

Now the rule separates math from judgment. When the receipt shows the tax, splitting it is arithmetic: the agent corrects it in the same run and reports what it changed. A choice of account or classification still goes to a person.

Three reasons, in the order we felt them.

The first was cost. Before an AI does anything, it reads everything it has been given: its instructions, the files, the conversation so far. That is called its context, and you pay for all of it on every run. A narrow helper that only needs one slice of the work has a small context, so it costs less. In our monthly review on October 7, 2026, a narrow helper started with about 9 pages of context. The general-purpose helpers in the same run started with about 100.* Smaller is better, as long as the helper still has what its task needs: it costs less, and there is less unrelated material for the AI to get mixed up by.

The second was reliability. A script gives the same answer every time. An AI told "never use this word" will still slip now and then, and you will hear about it from a client.

The third was review. When the AI only makes judgment calls, the person checking its work reads decisions, not arithmetic. That is where a reviewer's time is worth something.

We turn repeated mistakes into checks.

Our sales presentations are an example. The instructions said to explain every AI term the first time and avoid jargon. The AI still slipped jargon in, and I kept catching it in review. So the rule became a list: every word a reviewer rejected, next to its plain replacement. "Operational AI" became "we teach AI to work like your team." A checker now flags any word on the list before a presentation goes out, and each newly rejected word goes on the list. One sentence, our statement of what we believe, cannot change by a single word, or the check fails the presentation.

Once a month, we look back over past AI work for corrections that keep coming back. A correction counts only if it shows up in at least two separate places, so one bad afternoon doesn't become a rule. Each one gets the most reliable fix that can hold it, in this order:

1. An automatic check: a script or a test
2. A change to the skill's instructions
3. One line in the AI's standing instructions for the project
4. A note the AI keeps for next time
5. A to-do for a person

We only move down the list when the fix above can't do the job. Nothing changes without approval: the review ends with one decision for me to make.

Ask three questions about any AI your team uses today:

1. Which steps does a person still re-check by hand every time? Those are scripts waiting to be written.
2. Where is the procedure written down, and does the AI read the current version? If it lives in someone's head, the AI is guessing.
3. Which decisions would you never let the AI make alone? Those stay with a person, by design.

AI works best on narrow tasks, with only the access it needs, and with more freedom as it earns it. Writing down your procedure and turning its checks into scripts is most of the work. The judgment that is left is where your team's expertise lives, and that is the subject of part 2: why AI won't replace you.

* Measured in QAI's monthly review run on 2026-10-07: 5,800 tokens for the narrow helper's first request, against 66,000 to 72,000 for the general-purpose helpers. Tokens are the units AI providers bill by; we converted them to pages at about 0.75 words per token and 500 words per page.