Almost every prompt failure fits one of four classes. Each class has its own fix, so naming the class comes before editing.
Class | Symptom | Distinguishing test |
|---|---|---|
Missing context | The reply invents or requests facts the prompt never supplied. | Add the facts and rerun; a changed reply confirms it. |
Ambiguity | The reply is valid under a different reading of the prompt. | Define each key term once; the wrong reading disappears. |
Wrong format | The content is right but the shape is unusable. | Pin one example of the shape; the content stays and the shape fixes. |
Capability limit | The failure survives every phrasing of the same ask. | Split the task or verify the step outside the model. |
Context and ambiguity are the common classes. Format is the cheapest fix, and capability limits are the rarest verdict.
A tally over 20 failing prompts shows the pattern. The counts below are stated example parameters, and the arithmetic is exact.
triage tally over 20 failing prompts (example counts):
missing context: 9 -> 9 / 20 = 45%
ambiguity: 5 -> 5 / 20 = 25%
wrong format: 4 -> 4 / 20 = 20%
capability limit: 2 -> 2 / 20 = 10%
check: 9 + 5 + 4 + 2 = 20; 45 + 25 + 20 + 10 = 100%Fourteen of the 20 cases sit in context plus ambiguity. That is where diagnosis effort pays first.
Some asks exceed what a model can do reliably, and the boundary varies by model: measure it. Long inputs are less reliable than short ones.
Models tend to use information at the start and end of long inputs better than the middle. That is a tendency reported in studies, not a law.
No phrasing technique works on every model or every task. When three distinct phrasings fail the same step, treat the step as past the limit.
Missing context. Add the polarity and rerun; a changed reply confirms it.
Capability limit. Split the derivation into checked steps or verify it outside the model.