A model cannot ask you what you really meant before it answers — it answers the sentence you actually wrote. Clarity is removing the ambiguity that would force it to guess: one request per instruction, ordinary words, and every assumption stated rather than implied.
What clarity looks like
- One request at a time. A prompt that asks for a summary, a translation and a critique gets a shallow attempt at all three. Split them, or number them explicitly so each is answered in turn.
- Plain words. Jargon and internal shorthand ("make it pop", "the usual format") mean nothing outside your head. Say the concrete thing you want changed.
- No hidden assumptions. If the answer depends on a fact the model has not been told — which product, which version, which audience — it will invent one. State it instead.
- Unambiguous references. Avoid "it", "this" and "the above" when several things could match. Name the thing.
Common failures
- Compound questions. "Is this secure and how do I deploy it?" — the model optimises for whichever half it read as primary.
- Implied standards. "Make it professional" leaves the model to pick a register; "Rewrite in third person, no contractions" does not.
- Negation without an alternative. "Don't be verbose" tells the model what to avoid, not what to produce. Give a length instead.
Example
Vague
Fix the login bug and tell me if the rest of the code is okay.
Better
The login form rejects valid credentials when the email contains a plus sign. Find the cause in auth/validate.js and propose a fix. Answer only about that bug — I will ask about the rest of the file separately.