Robust prompt engineering

Robust prompt engineering is the practice of writing prompts that hold up under hostile, malformed, or simply weird input — not just the happy path you tested in a playground. Where basic prompt engineering optimizes for output quality, robust prompt engineering optimizes for consistency under pressure: the prompt behaves the same whether the input is a polite question, a 10,000-character paste, an emoji flood, or a deliberate attempt to derail it.

This is your first and cheapest layer of defense against the injection attacks covered in the previous topic. A fragile prompt leaks its instructions when asked, changes persona when a user role-plays, and treats retrieved documents as commands. A robust prompt won’t stop a determined attacker on its own, but it dramatically raises the cost of casual exploitation and reduces embarrassing everyday failures — hallucinated policies, broken output formats, off-topic rambling — that erode user trust just as badly as a breach.

Concretely, you’ll learn to put durable rules in the system prompt rather than the user turn, wrap untrusted content in delimiters like <user_input> tags and tell the model to treat it as data, restate critical constraints near the end of long contexts, and define explicit refusal behavior for out-of-scope requests. You’ll constrain outputs with schemas or structured outputs so downstream code never parses free text, and you’ll treat prompts like code: version them, write regression evals for known failure cases, and rerun the suite on every change and every model upgrade.