WWW’s core question
Is my prompt bad, or is the AI wrong?
Use WWW’s six verdicts to separate instruction problems, model failures, task limits, model choice, shared blame, and results that were actually correct.
Short answer
What is probably happening
Do not force every bad AI result into a two-sided blame game. A useful diagnosis has six possible outcomes: the instruction caused the miss, the model failed a clear instruction, both contributed, the model choice was wrong, the task itself was not answerable, or the result actually met the stated target.
Possible causes
The same symptom can have different verdicts
The diagnosis starts by reconstructing the intended result and freezing its requirements. Only then should you assign blame. If you decide who failed before defining success, the verdict will simply mirror your frustration.
Instruction issue
A required constraint, input, preservation rule, or success criterion was missing or materially ambiguous.
Model issue
The request was clear, feasible, and well-scoped, but the output violated it.
Both sides
The instruction left a genuine gap and the model handled that uncertainty irresponsibly.
Wrong model or tool
The request was sound, but the chosen mechanism was poorly suited to the required fidelity, context, access, or output type.
Task issue
Success required unavailable information, access, certainty, or a decision nobody had supplied.
Nothing went wrong
The output met the original requirements; disappointment revealed a new preference after the fact.
How to distinguish them
Check the evidence in this order
- 01
Write the intended result in one sentence without mentioning blame.
- 02
Separate explicit requirements from conservative inferences and later preferences.
- 03
Test each requirement against the output with concrete evidence.
- 04
Ask whether any model had the information, access, and mechanism needed to succeed.
- 05
Assign one primary cause only after the target and evidence are fixed.
WWW-style diagnosis
Example casenothing went wrong.
- Wanted
- A faithful five-bullet summary.
- Asked
- Exactly five concise bullets; keep caveats; add no recommendations or facts.
- Result
- Five faithful bullets with the caveats and no added claims.
- Why
- Wanting the result to feel more impressive appeared only after delivery. That is useful for the next instruction, but the original result met the frozen target.
What to do next
Fix the diagnosed cause
Repair the cause, not the symptom
Add missing constraints for an instruction issue; retry or switch workflow for a model issue; supply missing evidence for a task issue.
Keep the target stable
Do not change success criteria between the original and retry just to make one side look right.
Compare the retry requirement by requirement
A better-looking result can still regress on a requirement the previous round already solved.
When the instruction needs repair
A boundary-first template
I wanted: [intended outcome]. I explicitly required: [requirements]. I expected but did not state: [implicit expectations]. The AI produced: [relevant output]. The mismatch was: [specific differences]. Diagnose the primary cause as instruction, model, both, model choice, task, or no problem; cite the exact evidence before proposing a fix.
This template is relevant only when the evidence points to missing or conflicting boundaries. It cannot excuse a model that violated a clear instruction.
Questions
What people usually ask next
Is every misunderstanding a prompt problem?+
No. A model can misunderstand or ignore an adequate instruction, and some tasks fail because the tool or missing information makes success impossible.
Can both the user and AI be wrong?+
Yes. A vague permission can create uncertainty while the model still deserves blame for presenting guesses as facts or crossing reasonable boundaries.
Why include ‘nothing went wrong’?+
Because honest diagnosis must allow the evidence to clear both sides when the output met the stated target.