ChatGPTExample case
The page was replaced, not polished
The user named a visual destination but never protected the structure and copy that were meant to survive.
Mostly you
Case library
Bad result. Actual evidence.
Each case reconstructs the target, compares the instruction with the result, assigns one of six verdicts, and turns the failure into a useful next move.
ChatGPTExample case
A one-function fix deleted unrelated code
The edit scope was explicit; the model still replaced the file and dropped functions outside the requested change.
On the AI
ChatGPTExample case
The background changed—and so did the subject
The request required pixel-level preservation, while a generative editor redrew the whole image region.
Wrong model
ClaudeExample case
Claude rewrote the whole file
The user asked for one localized change and explicitly prohibited unrelated edits; the model returned a shortened replacement file.
On the AI
GrokExample case
The launch plan quietly invented facts
“Fill in the gaps” invited assumptions, but the model also failed to label those assumptions as invented.
Both
CopilotExample case
The policy became a summary
A preservation-heavy long-form rewrite was handed to a tool optimized for inline assistance and short transformations.
Wrong model
ClaudeExample case
The task was missing its source
The requested analysis depended on a document that was referenced but never attached or pasted.
Task blocked
ChatGPTExample case
The result was right; the target moved afterward
The output satisfied every stated requirement; “more impressive” appeared only after the result arrived.
Nothing wrong
WWW FOR IPHONE
Got a result that makes no sense?
Separate what you wanted, what you asked, and what the AI did. WWW turns the gap into a diagnosis.