Most prompts that break in production were not badly written - they were under-specified. They worked once, with one lucky example, then fell apart the moment real inputs showed up.
Specificity beats cleverness
The fix is rarely a cleverer sentence. It is saying out loud the things you assumed the model already knew: who is answering, what exactly to do, what facts to use, and what the answer should look like.
Tip
If you cannot name which part of the prompt a new instruction belongs to, it probably belongs in the context - or it does not belong in the prompt at all.
Build the habit, not the hack
Treat prompts like product copy: version them, test them against real inputs, and change one thing at a time. Reliability comes from process, not from a magic phrase.