AI fails in four repeatable ways. This one page names each failure and hands you the exact line to paste into any model so it stops.
Four Patterns,
One Page
The four ways AI output fails, and the line that fixes each
A flawed AI answer arrives in the same calm, confident voice as a correct one. You cannot see the seam. These are the four failures hiding under that voice.
It fills a blank the source left empty with something plausible.
A guess and a checked fact arrive sounding identical.
It validates whatever you give it, right when you needed pushback.
It takes ownership of a decision that belonged to you.
Read the AI's last answer, name what went wrong, and paste the matching line. Here are all four, in full.
It fabricates into empty fields, and the lie reads like the rest of the report.
If the source does not say, write UNKNOWN.
Do not fill a blank with something plausible.A guess in the same certain voice as evidence, so you act on tone.
Tag every claim HIGH, MEDIUM, LOW, or UNKNOWN,
each with its source. Never mark a claim HIGH
without a quote I can verify.Reflexive agreement, the AI going useless when you needed it to argue.
Do not be agreeable. Give me the three weakest
points in this, each with a specific reason. If it
is genuinely sound, say so plainly instead of
inventing objections.Overreach, the AI taking a decision that should stay with you.
Prepare everything I need to decide, then stop.
Do not make the call. End with:
My read (input, not the decision): [one line]
DECIDE: [ ]It is free. If it earns a spot on your desk, you can name a price to support it.
The page gives you the four moves. The Prompt Pattern Playbook teaches them until they are reflex.
Four moves you can use on a draft you are working on right now.