8. Checking Its Work Without Checking Everything

Here is the position you are now in. You have prompts that work, examples that transfer your voice, and a setup that remembers your context. You are going to start relying on this. Lesson 8 is the one that makes relying on it reasonable.

The good news is that the checking is cheap. It is maybe 60 seconds per piece of work that matters, and it is not "fact check everything", which is advice nobody has ever followed.

The finding this whole course rests on

In lesson 5 we asked it to explain mortgages, and it produced 4 wrong figures in one paragraph, including saying that a 15 year loan on $400,000 costs $20,000 in interest when the real figure is about $247,000. Off by a factor of 12. We asked in that same conversation what it was least confident about, and it did not mention the paragraph.

So we tried something else. We took the wrong paragraph, put it in a fresh conversation with no history, and asked one question:

Check the arithmetic in this paragraph. Show your
working for each figure and state whether it is right
or wrong.

"[paste the paragraph]"

It worked through the amortization formula, produced $2,661 a month for the 30 year and $3,591 for the 15 year, and marked all 4 of the original figures wrong. Its numbers came within 0.2% of ours. It closed with a line that is slightly funny given the circumstances:

The paragraph contains major arithmetic errors across all four figures. The author appears to have confused interest with total payments.

The author was itself.

Why that works, and it is the useful bit

Checking is a different task from producing, and it is much better at checking than it is at noticing while producing. While generating an explanation it is doing many things at once and holding a plausible shape. Given one paragraph and one job, it does that job properly.

It cannot reliably catch its own mistakes while making them. It can reliably catch them when you hand them back as a separate task.

The mechanics matter. Do it in a fresh conversation, and do not tell it the text is its own. In the original chat it is invested in what it already said and will defend it. With no history and no ownership, it just checks.

The 60 second routine

For anything you are going to send, publish, or act on:

1. Scan for numbers, names and dates. These are the failure category, per the table in lesson 5. Usually there are 1 or 2.
2. Ask "what are you least confident about?" in the same chat. Fast, catches the category even when it misses the instance.
3. Paste anything critical into a fresh chat and ask it to check. The trick above. This is the step that actually works.
4. Verify the 1 or 2 real specifics at the source. The price on the company's own page, the date on the actual document. Not everything. The 1 thing you would be embarrassed to be wrong about.

Steps 1 and 2 take 20 seconds. Step 3 takes 30. Step 4 is where the actual safety is, and it is 1 or 2 lookups, not a research project.

What to scale the checking to

Situation	How much to check
Understanding something for yourself	Nothing. Read it and move on.
Draft you will rewrite anyway	Nothing. You are the check.
Email going to a colleague	Read it. That is enough.
Anything with a number you will act on	Step 3 and step 4
Anything public with your name on it	All 4 steps
Money, health, legal, or someone else's job	All 4, then a human who knows

Most of what you do lands in the top 3 rows, which is why "check everything" is bad advice. It makes the tool feel exhausting and so people either stop using it or stop checking entirely, and the second one is worse.

Three more checks worth knowing

Make it argue against itself.

Give me the strongest case against the advice you
just gave. Then say which side you would actually
bet on and why.

If the opposing case turns out to be strong, you were handed a position dressed up as a fact. That happens more than you would expect on anything contested.

Ask what is missing rather than what is wrong.

What did you leave out of that answer that someone
in my situation would want to know?

Errors of omission are the ones you cannot spot by reading, because nothing looks wrong. Nothing is there.

Ask twice, separately. For a factual question that matters, ask it in a fresh conversation with different wording. Two answers that agree is weak evidence. Two that disagree is a clear signal to go and look it up yourself, and it costs 30 seconds.

The omission check, run for real

Errors of omission deserve their own demonstration, because they are the ones you cannot see. We gave it a piece of correct advice and asked what was missing:

Someone asked me how to safely lift a heavy box. I
told them: bend your knees, keep your back straight,
hold the box close to your body, and lift with your
legs.

What did I leave out that someone actually lifting a
heavy box would want to know?

Nothing in that advice is wrong. It is what everyone says. It came back with 6 things, and these 3 are the ones that matter:

• Test the weight first. Lift one corner before committing. If it is heavier than expected, stop.
• Clear your path. Know where you are going before you pick it up.
• Decide in advance to drop it. Better to let it go than to twist trying to save it.

Then it named the biggest gap: getting help. As it put it, that is the practical decision someone facing a heavy box makes first, before perfecting their form.

That is the shape of an omission. The original advice was accurate and it was also missing the first decision. You cannot catch that by proofreading, because there is nothing on the page to catch.

The two pass habit

The single change that made the biggest difference to how we use this: never accept a first draft as a final answer, and never check in the conversation that produced it.

1. Pass one, produce. Full context, examples, constraints. Iterate until you like it.
2. Pass two, check. New conversation. Paste the output with no history and no mention of where it came from. Ask it to find what is wrong or missing.

Splitting those 2 into separate conversations is what makes pass two work. The mortgage test above is the proof: same model, same question, and it defended the error in one context and demolished it in the other. The only difference was that the second one had no stake in it.

Four mistakes people make when checking

1. Asking "is this correct?" That is a leading question and you will usually get a yes. Ask what is wrong with it, or ask it to find the errors, which presupposes there are some.
2. Checking in the same chat. It will restate its reasoning rather than test it.
3. Telling it the text is yours or its own. Both bias the answer, in opposite directions. Give it the text with no story attached.
4. Checking everything, then nothing. The exhaustion pattern. Use the table above and check the rows that matter.

What we would tell someone finishing this course

Eight lessons in, here is the honest summary. This is a tool that is very good at language, structure and patience, and unreliable about specifics. Everything in this course is a way of leaning on the first list and defending against the second.

The people who get the most out of it are not the ones with clever prompts. They are the ones who supply real context, ask follow-up questions instead of accepting the first draft, and know which 2 things in an answer need checking. All 3 of those are habits, not knowledge, which means they only form by using it on real work.

Where to go now

Stop reading tutorials, including ours. Pick the 3 things in your week that are writing, reading, or explaining, and do those with it for a two weeks. That will teach you more than another 8 lessons would.

Then come back for the Beginner track, which picks up where this ends: longer documents, getting it to do multi-step work, and the point at which it is worth paying for.