Prompting and Working with Generative AI Tools
What to know
Effective prompting is specification. Strong prompts state role, task, context, constraints, format, and examples. Know techniques: zero-shot, few-shot (provide examples), chain-of-thought (ask for stepwise reasoning on suitable tasks), and role prompting. Understand iteration - evaluate output, refine the prompt, and keep a record of what worked.
Know the mechanics that shape results: context window limits, temperature (higher = more varied, lower = more consistent), and token limits. Understand RAG as grounding, and grounding/citation as a way to verify claims. Recognise hallucination and mitigation: ask for sources, constrain to provided documents, and verify facts independently.
Traps
Questions penalise vague prompts ("write something good") and reward explicit output formats. Beware assuming a confident tone means correctness, or that adding "be accurate" prevents hallucination. Another trap: pasting confidential data into a public tool - a privacy failure, not a prompting technique. Also note that chain-of-thought is not a universal fix and can introduce plausible-sounding errors.
Worked example
Instead of "Summarise this contract", use: "You are a contracts analyst. From the text below, extract termination notice period, renewal terms, and liability cap. Return a three-row table. If a field is absent, write 'not stated'. Quote the supporting sentence for each." Rule of thumb: specify the output shape and the behaviour when information is missing.