Your prompt isn't working. Pick your model, select the symptom, answer one question — get a specific fix with copy-paste snippets.
Which model?
What's the symptom?
Diagnostic question
Diagnosis
Fix — copy into your prompt
Model-specific note
Common combinations
Ignores instructions + Wrong format
These usually have one root cause: the prompt is unstructured. Group your instructions into labeled sections (Rules, Output Format, Task) so the model can't mistake one for another. For format: put a concrete example immediately before the final task instruction. Both issues often clear up together.
Makes things up + Inconsistent outputs
Both trace to underspecification. Set temperature=0, add an explicit context restriction ("only use the provided text"), and require a quote before each answer. Inconsistency drops sharply when the model is constrained to evidence rather than generation.
Loses track + Wrong length
Scope drift compounds length drift. Break the task into separate API calls with explicit per-call instructions (including length limits). If you give one call both a complex task AND a length constraint, the model will sacrifice the length constraint first.
Before you debug
Test with the simplest possible prompt first. Add complexity one piece at a time.
Change one variable per test. If you change three things and the output improves, you don't know which one helped.
Most "the model is bad" problems are actually "the prompt is ambiguous" problems.
What's obvious to you isn't obvious to the model. State the obvious.
If your system prompt is over 1000 words, it's probably too long. Prioritize ruthlessly.