Example: Putting the Bucket Together

The job

Let's go back to the invoices folder from the lesson on the loop, and this time ask for something more demanding: putting things in order. Renaming everything to a pattern, grouping by month, and producing a summary.

It's the kind of job where a misunderstanding causes real damage: if it renames two hundred files with the wrong criterion, you've got a bigger problem than the one you started with.

Move 1 — guardrails, before saying a word

In the same folder there's also a subfolder with the signed documents, which don't get touched. I don't ask it to leave them alone: I forbid it.

It's the difference we saw in the lesson on rules. If I write it into the request, I'm steering the model; if I set it as a prohibition, the tool blocks it, whatever the model decides. On something that must never happen, you go with the second one.

Same move: I open Claude Code in the invoices folder, not the folder that contains all of them. The first guardrail is always: don't let it into places it doesn't need.

Move 2 — the plan, before working

I switch to Plan mode and ask for the job. It goes and looks at how the folder is set up — how many files, what they're named now, what's inside — and comes back with a proposal.

And this is usually where the surprise shows up. In the plan I read something I hadn't said: for example that it would put files with no date in the name into a separate group, or that two invoices share the same number.

Those two details, without a plan, I would have discovered two hundred files in. Now I fix them with a sentence.

Move 3 — the yardstick, declared before you start

Before giving the go-ahead, I tell it how you know it went well:

• every file has a date and number in the name, in the same format;

• no file is left out of the groups;

• the number of files at the end is identical to the number at the start.

These are three conditions it can check on its own, while it works. The last one especially: it's the safety net against the classic disaster, a file lost along the way.

Move 4 — go, and what I watch

I leave the plan and let it work. From here my job is reading the few questions that remain — the ones I haven't already covered with rules — and looking at the proposed changes before accepting them.

If halfway through I notice it's heading somewhere I don't like, I stop it and say so. Correcting on the fly is part of the loop, not a failure.

The comparison: the other route

Now imagine the same job done by firing off the request and hoping.

The misunderstanding about the undated files would have happened just the same — only you'd have found out at the end. The signed documents would have been safe only as long as the model remembered your recommendation. And the only check on the result would have been you, by hand, on two hundred files.

The prep cost two minutes: one prohibition, one mode, three lines for the yardstick. And the rules stay there for next time too.

And now

First bucket closed: you know what you let it touch. But throughout this chapter you had to keep repeating how you work, what you want, how your stuff is organized.

In the next bucket we solve exactly that: what it knows about you, meaning how to get it to find things already written instead of telling it again every time.

Did you get it?

Two minutes, no grade. Trying to answer — even getting it wrong — cements the concepts far more than rereading. Not in the mood? Skip it: the lesson still counts as complete.

1In the example, what is the right order of moves?

Launch the job, then set boundaries if needed Guardrails and the right folder, then a plan, then go, with the yardstick declared before you start Plan, work, and I check everything myself at the end

2Why forbid the folder of signed documents, if you'd already said not to touch it?

Habit: two safeguards are better than one Because a spoken request steers the model, while a prohibition blocks it no matter what Because otherwise it can't see the other folders

3How much time does all this prep cost?

A couple of minutes, and you make it back the first time you avoid a misunderstanding Half a day: it's the price of working safely Nothing, because you set it up once and never touch it again