Chapter 5. Communicating with AI: Prompting and Iteration

5.1 The purpose of a prompt

A prompt is an instruction or set of inputs that guides an AI system. Effective prompting is less about finding magic words and more about defining the task, providing relevant context, specifying constraints, and stating what a good result should look like. Prompting works best as an iterative conversation: ask, inspect, clarify, and improve.

5.2 A practical prompt structure

A useful structure is: role or perspective; task; context; audience; constraints; output format; quality criteria; and verification instructions. Not every prompt needs every element, but complex tasks benefit from explicit requirements.

Example: 'Act as a project-management tutor. Explain the difference between a risk and an issue for a Level 5 IT student. Use a fictional CRM implementation as the example. Give a short comparison table, then three practice questions. Do not invent citations. Identify any assumptions.'

5.3 Break complex work into stages

Instead of requesting a complete business plan in one sentence, divide the task into stages: clarify the problem, identify the audience, list assumptions, draft an outline, develop each section, verify facts, review risks, and produce the final format. Staging makes errors easier to find and gives the user more control.

Ask the system to state what information is missing before it proceeds when missing facts could change the answer. For research work, request a source list and verify each source independently rather than accepting a plausible-looking bibliography.

5.4 Ask for evidence and uncertainty

Useful instructions include: 'Separate facts from assumptions'; 'Use only the supplied document'; 'Cite claims with sources that can be checked'; 'Tell me when the source does not support a conclusion'; and 'List the main uncertainties.' These instructions can improve discipline, but do not guarantee compliance or correctness.

For document analysis, identify the exact file and ask for page- or section-based evidence. For data analysis, ask the system to describe the variables, transformations, assumptions, and checks. Independently verify important calculations and outputs.

5.5 Iteration and evaluation

After the first response, provide specific feedback: what is missing, which audience it must serve, what is too long, which claims need support, and what format is required. Avoid vague requests such as 'make it better' when you can name the desired improvement.

Create reusable prompts for recurring tasks, but periodically test them. A prompt that worked with one model version may produce different results in another system. For important workflows, maintain a small set of test examples and compare output quality after changes.

5.6 Prompting mistakes

Common mistakes include vague objectives, too little context, conflicting instructions, asking for facts without requiring sources, supplying sensitive information unnecessarily, assuming that the first answer is final, and failing to check output against the real task. Long prompts are not automatically better; relevance and clarity matter more than length.