Prompt Engineering SIThe task is vague
Symptom: generic, safe answers. Fix: name the deliverable, the reader, and what it will be used for. “Summarize this” becomes “Summarize this for a manager who must decide by Friday.”
Context is missing or buried
Symptom: the answer ignores something important. Fix: put the facts that change the answer first, and label them, for example Background, Constraints, and Source text.
No output format
Symptom: an answer you cannot reuse. Fix: specify structure, such as a table with named columns or a fixed list of fields, plus an approximate length.
Instructions contradict each other
Symptom: inconsistent output between runs. Fix: read the prompt once as a stranger would, and remove repetition and requirements that pull in opposite directions, such as “be brief” and “cover everything.”
No rule for missing information
Symptom: confident guesses. Fix: tell the system what to do when something is absent, for example “Mark missing details as not specified.”
No check on the result
Symptom: plausible answers that turn out wrong. Fix: define a check in advance, and test on an ordinary case and an awkward one. The evaluation guide shows a small test set you can reuse.