The Reusable Prompt Skeleton

writing prompts for products

Part of: How to Build with AI: The Playbook

A prompt inside a product is not a chat message. It's a function's instructions. It runs a thousand times on inputs you'll never see, so it has to be precise and a little boring. Here's a skeleton you'll reuse in almost every project. System vs user Most APIs take two kinds of message: - System : the standing rules. Who the model is, what to do, what format to return. Set once. - User : the actual input for this run. Separate them and you write the rules once, then swap only the input each call. The skeleton Fill the braces per product. A summarizer's task is "summarize the text in 3 sentences." An extractor's Return is "a JSON object with keys name, email, phone." The rules that fix most bad output 1. Say the format exactly. "Return JSON" beats "give me the data." "Three bullet points" beats "be concise." 2. Show, don't just tell. One tiny example in the system prompt is worth a paragraph of rules. 3. Constrain the scope. "If the text has no email, return null" stops the model from inventing one. Why specific prompts win Vague prompts fail silently. The model returns something plausible that breaks your parsing step. A specific prompt is the cheapest reliability you'll ever buy. W