Build in the checkpoint

TL;DR

Verifying "when you feel like it" is a willpower-based system: you check once you're already worried, which is usually already late. Willpower-based systems are weak, because willpower runs dry exactly when you're busiest. The sturdier move is to wire a fixed checkpoint into the workflow — a default step that doesn't ask what kind of day you're having.

I used to think of myself as someone who verifies carefully.

And I did verify — when I felt I needed to. The problem is right there in those words. I checked when something felt off, when the output looked a little strange, when a hunch nudged me. Which means I only looked hard at the cases I already suspected.

So what about the times the agent got it wrong but looked completely normal? Those were exactly the times I didn't feel a need. So I didn't check. And those were exactly the times the error slipped through.

Turns out I wasn't someone who verifies carefully. I was someone who verifies once worried — a smoke detector that only goes off after the fire has caught.

01What you have to remember to do is what you'll someday forget

Here's the fracture point: "verify when you feel it" turns checking into a decision you make every single time. And every repeated decision draws from the same tank — willpower. That tank drains over the day, and it empties fastest on the day you're tired, rushed, with three other things shouting for attention.

Which means your verification is strongest on an ordinary day and weakest on a pressured one. And the pressured day is exactly when you push work out fast, skip steps, trust on faith. Your safety net thins out precisely when you need it thickest.

Think of a seatbelt. It works not because you're a careful person. It works because you don't re-decide every trip whether to wear it — you wear it, by default, even on the run to the corner. The moment it becomes a per-trip judgment call is the moment it starts to fail. The verify step is the same.

The piece "'Done!' is a claim" in the shipping cluster is about the psychology — why you're so quick to believe the agent's "done." This piece is about the layer beneath that: even once you know not to trust it, you still need a check that doesn't depend on whether you happened to remember to distrust it today.

✕ Checking on a hunch

Frequency: high when you're free, bottoms out on the busy day

▼

Catch: only catches what you already suspected — misses what looks normal

▼

Mental load: burns energy weighing "do I need to check this one"

✓ A fixed checkpoint

Frequency: steady, because it's tied to the workflow, not the mood

▼

Catch: catches what you didn't expect, because you check even with no suspicion

▼

Mental load: nothing to decide — the step always runs, so your head is freer

Look at the middle row: the biggest difference isn't that you check more. It's that you check even when there's no reason to — and that's exactly where silent breakage chooses to hide.

02Signs your checking isn't a system yet

Most people believe they verify more regularly than they do. A few signs it's still running on willpower rather than mechanism:

Your checking still runs on willpower when…

The busy day is when you skip itif the verify step is the first thing to drop under pressure, it's an option, not a default

You only check what you already suspectif you can't recall the last time you checked something that looked perfectly fine, the fine-looking ones are going straight through

"I'll check it later" — and later never arrivesdeferring a check is skipping it on a timer, and that timer rarely rings

You can't name this workflow's check stepif you have to think to answer "how do I verify this kind of work," there is no step yet — just a vague habit

Every sign above shares one denominator: checking is something you remember to do, not something that just happens. What you have to remember will slip — not today, but on the very day you need it most.

03Wire it in; don't hand it to your memory

The way out isn't resolving to check harder — resolve draws from the same tank that's going to run dry. The way out is to make checking a fixed step of the workflow itself, so there's nothing left to decide.

Concretely: for each kind of work you regularly hand the agent, define one non-negotiable check ahead of time — and wire it into the request at the start, not the end. "When you're done, run it and paste me the real output." "After you fix it, list every place you touched." That step doesn't change with the day, doesn't change with your mood, doesn't need you to remember it — it's part of the work, the way the seatbelt is part of the driving.

The nice part is that once the check lives inside the workflow, it stops draining you. You no longer weigh "do I need it this time" — that question disappears. The safety net stays evenly thick, even on your most exhausted day, because it no longer depends on you being sharp enough to remember to stretch it out.

The right question isn't "should I check today." It's "what is this work's check step, and is it already sitting inside the workflow." Answer the second, and you never have to ask the first again.