Edge Cases & Defensive Prompting

80% of production prompt failures come from inputs you never anticipated.

The goal isn't to build a perfect prompt for ideal conditions. It's to build a prompt that fails predictably on the inputs you didn't design for, and refuses cleanly on the inputs designed to break it.

Three categories

Input edge cases

Bad data, not bad intent.

• Empty input — user submits whitespace or nothing.
• Excessive length — input exceeds context budget for the rest of the prompt.
• Special characters — control characters, zero-width spaces, mixed Unicode.
• Multilingual / mixed languages — single message in two languages.
• Malformed text — truncated JSON, broken markup, OCR garble.
• Ambiguous input — multiple valid interpretations.
• Contradictory input — user states A then states not-A.