Checking Work You Did Not Watch

The genuinely hard problem

Everything up to here has been preparation. This is the difficult part, and it is difficult in a way that does not go away with practice.

You are being asked to judge a piece of work whose making you did not see. You have the result and nothing else. You did not watch which file it opened first, or the moment it found two conflicting dates and picked one, or the point where it could not find something and carried on regardless. All of that is invisible to you, and the finished document is smooth, confident and gives you nothing to catch on.

Claude asks for your review and approval before shipping the work. That is the safeguard the whole product is built around. It is worth exactly as much as the quality of the review you give it, and no more.

In plain English

Run report:: The account of what a job actually did, as distinct from what it produced.
Assumption:: A gap the work filled in by itself. The most important thing it can tell you.
Silent gap:: Missing information that was smoothed over instead of flagged, leaving no trace in the output.
Sampling by consequence:: Checking the few items that matter most in full, rather than skimming everything equally.
Spot check:: Tracing one specific claim in the output all the way back to the file it came from.

What a trustworthy run report contains

Our agents course establishes the three questions to put to any run: what did it read, what did it change, what did it send. Keep those. For work you did not watch, add a third layer, and this is the one that does the real work.

What it did. The steps taken, in order, and the files touched. This tells you whether the job it ran is the job you thought you briefed. It is surprisingly often not.

What it could not find. Named, specifically. Not a general note that some information was unavailable, but which file was missing, which field was empty, which item had no status. This is the difference between a gap you can chase and a gap you never learn about.

What it assumed. The most valuable section and the one you have to ask for explicitly. Every judgement call made in your absence: which of two conflicting figures was used, how an ambiguous date was interpreted, what a blank field was taken to mean. A run that reports no assumptions on a real-world job has almost certainly made several and not noticed.

The run report to insist on

WHAT I DID
step 1  read 14 files in Northgate/Q2
step 2  extracted status and date from each
step 3  compared against last month's summary
step 4  wrote northgate-q2-summary.md
WHAT I COULD NOT FIND
no status field in three files, named below
no June entry in the tracker at all
WHAT I ASSUMED
two files gave different totals; I used the more recent file
one date was written 03/04; I read it as day first
blank owner fields were treated as unassigned rather than as errors
WHAT I CHANGED   created one new file, nothing else
WHAT I SENT      nothing
STATUS           waiting for your approval

Checkpoint

What it did, what it could not find, what it assumed. The third one is the one you have to ask for, and it is where the errors are.

Where the errors actually hide

Not where you look. Reviewers naturally check the beginning and the end of a document, so those tend to be fine.

The confident middle. Paragraph four of seven, where attention has dropped and the prose has settled into a rhythm. Errors survive there.

Transcribed numbers. A figure that was correct in the source and is now one digit different. There is nothing in the sentence to signal it, which is exactly why it survives review.

The joins between sources. When material from two files is combined, the seam is where a date from one gets attached to an item from the other. Any claim built from more than one document deserves more suspicion than one built from a single document.

Silent gaps. Something missing was smoothed over rather than flagged, so there is no hole to notice. This is what the gap rule from lesson three is for, and it is why a job briefed without one is close to unreviewable.

The thing that was not there at all. The hardest of the lot. An item that should appear in the output and does not, because the source never mentioned it. Nothing in the document can prompt you to notice an absence. Only your own knowledge of the work can.

❌ Weak prompt

Prompt

Looks good, approve.

Output

The summary ships. Three weeks later somebody notices the March figure was carried over from February, because the March file was missing and nothing said so.

Approval given to the appearance of the work rather than to the work. A smooth document is evidence of nothing at all, and this is the failure mode the cloud makes easiest.

✅ Good prompt

Prompt

Before I approve: list every assumption you made and every place you could not find what you needed, each one naming the specific file or field. Then show me, for the three largest figures in the summary, the exact file and line each came from.

Output

Four assumptions listed, two missing files named, and three figures traced back to source. One of the three came from a file the reviewer knew was superseded.

Two minutes of work that catches the class of error a full read-through would have sailed past, because it interrogates the process rather than the prose.

Sample by consequence, not by proportion

You cannot check everything, and pretending otherwise produces the worst outcome of all: a review that feels thorough, takes twenty minutes and catches nothing.

So stop distributing your attention evenly. Skimming all forty rows gives every row the same two seconds, which is not enough to verify any of them. Instead, find the three items where being wrong costs the most, and trace those all the way back to source. Full check, no skimming. The remaining thirty-seven get a glance for anything obviously odd.

The review pass, in order

1  READ THE ASSUMPTIONS FIRST, before the output itself.
You are looking for a judgement you would not have made.
2  READ WHAT IT COULD NOT FIND.
Ask whether any absence changes the conclusion.
3  IDENTIFY THE THREE HIGHEST CONSEQUENCE ITEMS.
The figure someone acts on. The date someone commits to. The name.
4  TRACE THOSE THREE TO SOURCE. Open the file. Compare the characters.
5  ASK WHAT IS MISSING that no one would notice was missing.
6  ONLY THEN APPROVE.

Ask this before approving anything

Before I approve, answer these separately from the output:
Which claims here rest on more than one source, and which sources?
Which numbers were copied, and which were calculated by you?
What did you leave out, and why?
If exactly one thing in this is wrong, what is it most likely to be?
Quote the file and line for each figure I name below.

The cloud makes the approval step easier to give badly. Reviewing on a phone, hours later, between other things, is the situation in which everything above quietly stops happening. If a job's output carries real consequence, do not approve it from a corridor. Let it wait until you are at a desk with the source files open.

📝 Quiz

Question 1 of 4

Which part of a run report does the lesson call the most valuable and the easiest to omit?