When an answer disappoints, look at the request before blaming the model. These five fixes cover most of what we see.
1. Say what you want made, first
"Draft a two-paragraph reply to this complaint" beats "Can you help with an email?". The model plans around the first thing you ask for. Put the output, and its length, in the opening sentence.
2. Name the reader
A summary for a deputy secretary and a summary for the front counter are different documents. One line does it.
This is for {reader}, who needs to {decision or action}.
3. Paste the source, don't describe it
"Our leave policy says something about carry-over" invites a guess at what a typical leave policy says. Paste or attach the actual clause. If it's long, paste the relevant section and say which part matters.
This is also where most invented detail comes from. A model asked about material it hasn't seen will fill the gap with something plausible.
4. Fix with a follow-up, not a fresh start
The first draft is rarely the last, and that's fine. Tell it exactly what to change and what to leave alone:
Make the second paragraph half the length and more direct. Keep everything else.
Starting a new chat throws away the context you already built. Rewriting the whole prompt usually loses something that was working.
5. Check the parts that would hurt if wrong
You don't need to verify every sentence. You do need to verify names, dates, dollar amounts, clause numbers and anything you'll quote to someone outside your team. Those are the details a fluent answer can get wrong while sounding completely sure of itself.
For questions where a wrong answer costs real money or trust, turn on Consensus. Two independent models answer, and the Second opinion section shows you where they differ, which is usually exactly the point worth checking.
Put this into practice with K9 AI.
Australian-hosted AI with redaction, classification and audit built in. Try it free for 14 days with no credit card, or talk to us about a pilot.