Recognising it is not working.
THE SIGNS
Checking takes as long as doing Errors keep reaching output The output needs rewriting rather than editing Verification is being skipped because of volume Nobody can explain what the system decided
THAT FOURTH ONE
The most dangerous. When review becomes nominal, the control has failed even though it appears to exist.
WHAT TO DO
Stop using it for that task Continue for the tasks where it works
Partial adoption is a legitimate outcome.
WHAT TO CHECK FIRST
Whether the problem is the prompting rather than the tool.
Most disappointing output comes from insufficient context.
WHEN TO STOP ENTIRELY
Rarely. Usually specific tasks are unsuitable while others are fine.
WHAT NOT TO DO
Continue because you invested in it Continue because stopping looks like failure
THE HONEST QUESTION
Is this producing better work, or just faster work with more errors?
