Claude CoworkClaude Cowork · Pro, Max, Team and Enterprise plans
1
Curious builder0 XP earned · 300 to level 2
0 daysFinish a lesson to begin
Badge collection0 of 6 unlocked
21 small wins to finish your pathNext lesson

When it goes wrong

A twenty minute task that produced the wrong deck is the normal case, not the failure case. What you do next decides whether it happens again.

Read the run before you rewrite anything. The plan it made, the steps it took and the files it produced together tell you which of the five things in lesson 4 was missing. Almost every bad result traces back to one of them.

Four failures, and where they come from

What you gotWhat was missing
It used the wrong filesInput. The folder was named loosely, or it found an older copy
The numbers are right, the document is uselessAudience. It did not know who reads this
It quietly dropped rowsEdge cases. The mess was not described, so it decided
It invented a figureThe stop-and-ask sentence was not there

Notice what is not on that list: the model being bad at the job. It happens, but it is far down the list of causes, and rewriting the brief is cheaper than assuming it.

Fix the brief, not the output

The temptation after a bad run is to ask for corrections in the same session until the result is right. That works once and teaches you nothing. The result you want is a brief that would have produced the right thing the first time, because next month you are going to run it again.

markdown
Not: "no, put revenue in thousands"

But, in the brief: "Revenue in thousands, rounded to one decimal,
with the unit stated in the column header."

When the answer is a smaller task

If two rewrites have not fixed it, the job is probably two jobs. Split it: clean the data and check the result, then produce the summary from the cleaned file. Each half is easier to describe and much easier to verify.

Do not delete the evidence
Keep the bad run. It is the clearest record you will ever have of what your brief actually said, as opposed to what you meant.
Try it yourself
  • Find a run of yours that went wrong and name which of the five it was missing.
  • Rewrite the brief so that failure is impossible, then run it on the same folder.

Little by little, you're building something great.