Precision is the editing pass on your own prompt. Every vague quantifier, unbounded scope and unstated exclusion is a decision handed to the model — usually to be resolved as a longer, hedged, more generic answer than you wanted.
Tighten
- Replace quantifiers with numbers. "A few examples" becomes "exactly three examples"; "short" becomes "under 120 words".
- Bound the scope. Name the files, the date range, the sections in play — and say that everything else is out of scope.
- State exclusions. "No preamble", "no apologies", "do not restate the question", "do not suggest alternatives I did not ask for".
- Fix the vocabulary. If a term is ambiguous in your domain, define it inline the first time you use it.
Guardrails
- Handle the unknown case. Say what to do when the answer is not in the material: "reply exactly UNKNOWN" beats an invented answer.
- Constrain the output space. Enumerate the allowed values for a classification instead of accepting free text.
- Beware over-constraining. Too many rules crowd out the task; keep the constraints that change the answer and drop the decorative ones.
Example
Vague
Summarise this report and give some recommendations.
Better
Summarise the attached report in exactly 5 bullets, each under 20 words. Then give 3 recommendations, ranked by expected impact, each with the sentence from the report that supports it. Use only the report; if something is not covered, write NOT IN REPORT. No introduction or closing remarks.