prompt-creation

Follow the 8-step workflow: define the objective in one sentence, identify the audience, list constraints, choose techniques like chain-of-thought or few-shot, write using the 9-element structure, add guardrails, test mentally against edge cases, and iterate.

It covers task context, tone context, task description with rules, few-shot examples, input data, immediate task, precognition scratchpad, output formatting, and assistant prefill. Each element addresses a distinct aspect of prompt behavior.

Two to five examples are sufficient; more examples do not improve results. Prioritize diversity over quantity, include edge cases, and keep formatting consistent across all examples.

Use explicit grounding instructions like answering only from provided context, add a confidence fallback phrase, ask the model to quote relevant passages before answering, and use temperature 0 for factual retrieval tasks.

XML tags work well for structured instructions with nested rules, identities, and output specifications. Markdown headers suit simpler section separation. Both clearly delimit instructions, context, task, and expected format.

Text below roughly 12px may be misread, handwriting accuracy varies, and dense diagrams over 20 nodes should be split into sections. Keep images between 768px and 2048px, since larger sizes add cost without quality gains.