Story audio is generated using AI
This is the eighth instalment of AI Fluency Corner, a 16-part weekly series in Business Day building one connected mental model of AI, in plain language.You have hired the sharpest analyst in history — one who has read every book, every report, every regulation ever published, is available at midnight, charges nothing and never complains. There is one catch: they are catastrophically literal. Tell them to write something on climate and they produce something on climate ― technically responsive, entirely generic and useless. Give them a precise brief and they are extraordinary. That catch has a name. It is “prompting”. Closing that gap is the difference between AI as a party trick and AI as genuine leverage.A 2023 McKinsey study confirmed it: the highest AI productivity gains went not to those with the most powerful tools, but to those whose teams had learned to instruct them well. Nobody taught that skill. This installment does.A prompt is not a question. It is a brief.A large language model does not read your mind. It reads your brief — words, context, implied assumptions and silence — then generates the most statistically probable continuation. Silence is not neutral. It is permission to guess, using the most average, most generic version of what you might have meant. A strong prompt contains six ingredients: task, context, role, source, format and quality standard. Here is what those six ingredients do to the same request:WeakYou: “Summarise this report.”Model returns five generic paragraphs. Correct register. Wrong emphasis. Could apply to any report ever written.StrongYou: “Act as a senior project manager. Summarise this report for a director with five minutes before a board meeting. Focus on decisions required, risks, budget impact, unresolved issues. Five bullets. Do not add information not in the report. Flag anything uncertain.”Five decision-ready bullets. One data gap flagged. Twenty minutes of rework eliminated.The prompt formula:Act as [role]. Using [source], produce [output] for [audience].Format as [structure]. Prioritise [criteria]. Avoid [failure modes]. Flag assumptions.Nobody would ask a junior analyst to “do something with this file” and expect a board-ready output. Yet sensible professionals do precisely that with AI every working day. That structure, applied consistently, changes what you receive. Not marginally. Categorically.When AI gets it wrong: hallucinations, stale data and the wrong toolHallucinations are the AI problem most discussed and least understood. The model was not trained to say “I don’t know”. It was trained to produce something fluent. Without grounding, it invents plausible statistics and clauses that dissolve the moment someone asks where they came from — at exactly the worst moment.Three techniques address this: ground the model in your evidence and constrain it explicitly; use chain-of-thought prompting (“explain your reasoning step by step before answering”) to surface errors before they reach your output; and instruct it to declare uncertainty (“If you are not certain, say so”). One sentence. Enormous difference.High-stakes prompt: grounding + verificationYou: “Using only the attached document, answer the question below. If the answer is not in the document, say so — do not supplement from general knowledge. Structure your response as: (1) Confirmed by the source. (2) Inferred. (3) Must be independently verified before action.”Related but distinct: the data currency trap. Most models have a training cutoff — no knowledge of recent legislation, this quarter’s results or last week’s amended regulation. For time-sensitive work, use a model with live web access or add a verification flag. And the wrong-tool trap: ChatGPT, Claude, Copilot and Gemini are language systems – excellent for drafting and reasoning, not for live data or legal search. Match the instrument to the task.The model is the instrument. The prompt is the score. A poorly written score produces noise, even from a world-class orchestra.The memory problem — and how to work around itAI has no memory between sessions. Every conversation starts blank. Within a session, the context window limits how much the model holds in active consideration — exceed it and it silently drops earlier content, forgetting your instructions before it forgets the document. Workarounds: open every session with a two-sentence context brief. For ongoing projects, maintain a “project summary block” and paste it at the top of each new thread. When a conversation turns contradictory, start fresh — continuity built on noise is drift, not continuity.Session reset prompt:You: “Ignore all prior context in this thread. The current brief follows. [Paste project summary block here.] Treat this as the only controlling instruction for everything that follows.”Use the ecosystem intelligently: Copilot is already embedded in Word, Excel, PowerPoint, Outlook and Teams. Claude and ChatGPT run in browsers and on mobile. Run them in parallel — draft in Word with Copilot, analyse in Claude, stress-test in ChatGPT — and bridge sessions by pasting relevant excerpts into each new thread. A reusable prompt template pasted fresh each session is not a workaround. It is professional practice.Parameters, precision and the three questions before you sendTemperature governs creativity versus precision – high for brainstorming, low for anything that must be accurate before it is interesting. Specify output length (“Respond in no more than 150 words” is instruction, not suggestion) and use examples over adjectives: paste a sample paragraph and write “match this register” rather than “make it professional”, which can mean anything from measured to written by a committee that has never met a reader.Before any prompt, run three questions: what context am I providing? What exactly am I asking and how will I know it was done well? What must it not do, assume or invent? These are the questions any manager asks before delegating. The model is not the bottleneck. Vague delegation is.South Africa’s AI dividend belongs to people who turn messy intention into clear instruction and fluent output into verified, usable work. ONE TASK THIS WEEKTake one prompt you actually use at work — a summary request, a draft, a planning task — and rebuild it using: role + task + context + source + format + quality standard + verification instruction. Run both versions. The difference will be obvious. Not because the AI became smarter. Because you did.Next week: Multimodal AI — what it means that modern systems can now see, hear and generate images and video and how these capabilities are already reshaping work far beyond text.
