Prompting as a Product Skill

Large language models are next-token predictors trained on massive text datasets. They don't 'think' in the human sense — they predict the most likely continuation of the text you provide. Understanding this changes how you write prompts. The model is pattern-matching against everything it's seen; your prompt is the setup that determines what patterns it draws on. This has practical implications. 'Be helpful' in a system prompt activates different patterns than 'You are a senior occupational therapist writing clinical documentation for NDIS reports.' The more specific and contextually grounded your prompt, the more consistent and useful the output. Vague prompts produce average outputs — the mean of the training data. Specific, well-framed prompts produce targeted, high-quality outputs. The prompt stack has three layers: System Prompt (the persistent context that frames all interactions — who the model is, what it's doing, what format it should use), User Message (the immediate input or request), and Context (relevant data provided alongside the request). Each layer contributes to the output, and each can be optimised independently.

Few-shot examples are the most reliably effective prompting technique. Instead of explaining what you want, show the model three examples of input→output pairs, then provide the new input. The model generalises from the examples. For a BathCheck feature that generates professional assessment language from structured data, provide three examples of raw assessment data and the desired professional output — the model learns the style, terminology, and structure from those examples. Chain-of-thought prompting improves accuracy on reasoning tasks. Adding 'Think step by step' or 'Reason through this carefully before answering' to a prompt activates a more deliberate reasoning pattern. For complex assessments, risk scores, or structured analysis, this technique measurably improves output quality. System prompt construction: the best system prompts have four components: (1) Role and expertise — who the model is and what it knows. (2) Task specification — exactly what it should do. (3) Output format — JSON, markdown, specific structure. (4) Constraints — what it should NOT do, edge cases, failure modes. Investing 30 minutes in a well-crafted system prompt saves hours of output correction downstream.