AI OUTPUT DEBUGGER
WWW (What Went Wrong) is an AI output debugger that determines whether a bad AI result came from the user’s instruction, the model, the task, the model choice, both sides, or whether nothing actually went wrong.
“Premium” describes a direction, not the boundaries of the edit. The request did not say that the layout, sections, copy, and behavior were preservation constraints, so a redesign was a plausible interpretation.
Your instructionMissing preservation constraints
AI executionOver-helpful, but defensible
Model choiceFine for this job
THE THING BETWEEN PROMPTING AND BLAME
WWW reconstructs what you meant, freezes the requirements, and compares them with what the AI actually produced. It does not start by assuming the user wrote a bad instruction—or that the model is innocent.
A missing boundary, requirement, or success criterion made the miss defensible.
The instruction was adequate and the model still crossed a clear boundary.
The instruction left a real gap and the model handled that gap badly.
The request was reasonable, but the selected model or tool was a poor fit for the job.
The job required missing information, unavailable access, or a decision only the user could make.
The result met the stated target; the disappointment came from a preference added afterward.
ONE CASE, PROPERLY DEBUGGED
The same output can be a model failure in one case and a reasonable interpretation in another. The exact ask decides which.
“Make this landing page look more premium.”
A new layout with a large gradient hero, rewritten copy, two removed sections, and a different information hierarchy.
The changed elements were never named as protected. The model did too much, but the original instruction gave it permission to decide what “premium” meant.
EXAMPLE CASE LIBRARY
Browse complete examples with the original goal, exact instruction, observable result, WWW verdict, evidence, fix, and general lesson.
The user named a visual destination but never protected the structure and copy that were meant to survive.
The edit scope was explicit; the model still replaced the file and dropped functions outside the requested change.
The request required pixel-level preservation, while a generative editor redrew the whole image region.
HOW WWW WORKS
Separate explicit requirements, conservative inferences, and any clarification that would materially change the call.
Compare the instruction, result, task conditions, and model fit against the same frozen success criteria.
Repair the instruction, retry unchanged, switch tools, supply missing evidence, narrow the task—or keep the result.
START WITH THE FAILURE YOU SAW
A practical way to tell whether ChatGPT ignored a clear instruction, missed an implicit constraint, followed a conflicting instruction, or lost the task in too much context.
Read the diagnosisRecover safely, inspect the diff, and determine whether ChatGPT lacked an edit boundary or violated one that was already explicit.
Read the diagnosisDistinguish an underspecified edit from over-helpful model behavior, then write preservation constraints that make unwanted changes testable.
Read the diagnosisDetermine whether Claude was invited to refactor, lost necessary context, or violated an explicit one-function edit boundary.
Read the diagnosisWWW FOR IPHONE
Bring the exact instruction and output. WWW reconstructs the target, checks the evidence, assigns the cause, and gives you the next move.