The Boardroom Coach
Bite-sized tutorials for the six leading AI models · Issue 1 · Week of 2 August 2026
Reading time: about 8 minutes
The single skill that changes everything: writing the ask well.
You do not need to be technical to get far more out of an AI model — you need to give clear instructions. Every one of the six leading labs publishes the same core advice, because a well-built prompt is the cheapest performance upgrade there is. It needs no new software, no budget and no engineer. This issue teaches the fundamentals once, then shows you the official guide for the exact tool you use — so a vague request becomes a reliable one.
Most people talk to an AI the way they'd fire a quick question across the office: short, assumed context, hope for the best. The models are good enough that this often works — and just unreliable enough that it quietly wastes your time. Prompt engineering is simply the discipline of removing the guesswork. Get it right and you turn a clever-but-erratic assistant into something closer to a well-briefed colleague. This week we cover the fundamentals that every model-maker agrees on, and point you to each company's own free tutorial.
Whether you use Claude, ChatGPT, Gemini or any other model, the same six habits carry most of the improvement. Read them once and you will spot the gaps in almost any prompt.
Start by telling the model who it should be. "You are a cautious financial analyst" or "You are a plain-English copy editor" sets the tone, vocabulary and level of detail before you even state the task.
State exactly what you want done, for whom, and to what end. "Summarise" is weak; "Summarise this contract's payment terms for a non-lawyer, flagging anything unusual" is a brief.
Paste the document, the data, the background. Models reason far better over text you provide than over what they half-remember. If the answer lives in a file, give it the file.
Ask for the shape you want: a five-bullet summary, a table with three columns, a 100-word paragraph, valid JSON. Naming the format removes a whole round of "not quite — can you redo it as…".
One good example of input-and-desired-output (a "few-shot" prompt) is worth a paragraph of instruction. It shows the model your standard rather than describing it.
For anything with reasoning, add "think step by step before answering". Then treat the first reply as a draft — refine your prompt based on what came back. Prompting is a conversation, not a single shot.
Here is a bare request rebuilt with all six habits. Copy it, swap in your own material, and feel the difference:
You are an experienced management consultant writing for a busy executive. Task: Read the meeting notes below and produce an action summary. Notes: """ [paste your meeting notes here] """ Format your answer as: - 3-sentence overview - A table with columns: Action | Owner | Deadline - One "risks to watch" bullet list (max 3 items) Think through the notes step by step before writing, and only include actions that were actually agreed — do not invent owners or dates.
Notice there is nothing technical here. It is just a clear brief — role, task, context, format, a constraint against making things up. That is prompt engineering.
The fundamentals are universal, but each lab documents them in its own voice — and each has a free official guide worth bookmarking for the tool you actually use.
Anthropic's guide leans hard on being explicit, using XML tags (like the """ block above, but with tags such as <notes>) to separate instructions from content, and giving Claude room to reason. Their free interactive tutorial walks you through it chapter by chapter.
Official: Prompt engineering overview · Interactive Prompt Engineering Tutorial (GitHub)
OpenAI distils prompting into six strategies: write clear instructions, provide reference text, split complex tasks, give the model time to think, use tools, and test systematically. The Help Center version is written for everyday ChatGPT users; the API guide goes deeper for builders.
Official: Best practices for ChatGPT · Prompt engineering (API docs)
Google's "prompt design strategies" page is refreshingly practical: give clear constraints, specify the response format, add examples, and prototype freeform, structured and chat prompts in Google AI Studio. A strong first stop if Gemini is your default.
Official: Prompt design strategies (ai.google.dev) · What is prompt engineering (Google Cloud)
Meta's how-to guide covers zero-shot and few-shot prompting, role-based prompts and techniques to reduce hallucinations with clear context and examples. The Llama Cookbook adds runnable notebooks if you want to try it in code.
Official: Prompt engineering (llama.com) · Llama Cookbook (GitHub)
Microsoft's small Phi models reward structured prompting. The Phi Cookbook shows a technique called "Guidance" that steers the model token by token for highly structured outputs; Microsoft Learn's general prompt-engineering page covers the fundamentals for Foundry and Azure users.
Official: Phi Cookbook (GitHub) · Prompt engineering techniques (Microsoft Learn)
Amazon's Nova guide is explicit about the three roles — system, user and assistant — and recommends breaking complex jobs into chained steps. It also has task-specific pages (text, vision, image and video generation) once you move beyond the basics.
Official: Text prompting best practices · General prompting tips (Nova user guide)
Take a task you did badly with AI recently — a summary that missed the point, an email that sounded wrong. Open your model of choice and rewrite the request using all six habits: role, specific task, context, format, an example, and a "think first" instruction.
Run the old prompt and the new one side by side. Keep the better wording as a reusable template — a short library of good prompts is the highest-return thing a non-technical user can build this year.
Claude — Anthropic's Claude Opus 5 (released 24 July) held the top of the Artificial Analysis intelligence index through the week. anthropic.com/news
ChatGPT / GPT — OpenAI cut GPT-5.6 API prices on 30 July, dropping the "Luna" tier by 80% as competition shifts toward cost. openai.com
Gemini — Google shipped Gemini 3.6 Flash (21 July) as its new default: 1M-token context, cheaper per token and up to ~17% fewer tokens used. Gemini API release notes
Llama — No major model release this week; Meta's Build with Llama cookbook and how-to guides remain the reference for Llama 4 prompting. developer.meta.com
Phi — No major update this week; Microsoft's most recent addition is Phi-4-reasoning-vision, a compact multimodal reasoning model documented by Microsoft Research. microsoft.com/research
Nova — No major update this week; Amazon Nova 2 (Lite generally available, Pro in preview) is the current generation, with adjustable low/medium/high "thinking" intensity. aws.amazon.com
Every link below is an official, free learning resource. This is your reference library for the week — and the standing hubs are worth a permanent bookmark.
For educational information only — model features and pricing change; always check the official docs.