> ## Content Index > Fetch the complete content index at: https://bayesiantheory.com/llms.txt > Use this file to discover other available public pages before exploring further. # AI Bayesics #2: Communicating effectively with AI - URL: https://bayesiantheory.com/ai-bayesics-2-communicating-effectively/ - Published: 2026-03-13T12:45:11.000Z - Updated: 2026-10-06T11:37:39.000Z - Description: Sick of generic AI outputs which miss the mark? It's as easy as Specify, Modify, Verify. - Author: Gregg Bayes-Brown - Tags: AI Bayesics, Writing, Skills, Training, AI, Coaching, Writing with AI, AI enablement, AI training, Newsletter, #Migrated-1791211413550, #Import 2026-10-05 14:44 You’re at your desk innocently doomscrolling Bayesian Theory when the email hits: “Hi whatsyourname, Can you do the thing? Make no mistakes. Kthxbi, Boss” You freeze. The thing? Do something? No mistakes? In a panic, you crack open a tin of word soup and apply it liberally to M365\. You click send. The anxiety fades until you get another a ding: a meeting with your boss. At the end of the day. With HR. Despite “excellent communication skills” appearing on every job ad since Ugg hired Ogg to paint his cave, few professionals actually possess them. Unfortunately, many believe they do. In reality though, slapping two sentences together does not make you Shakespeare. This same misplaced overconfidence carries over into AI use. Nearly all AI users grossly overestimate their skill by a staggering degree: research suggests self-reported proficiency vs actual quality is functionally zero. Why? Because effective AI directly proportional to your ability to communicate. [Subscribe now](#/portal/signup) ## Specify Let’s play a game of good prompt, bad prompt. Put this into your AI of choice: Write me a blog about AI risk. What you’ll likely receive will be the epitome of AI slop. Generic. Bland. An Ed Sheeran song in prompt form. Many users are unaware of the professional risk to their credibility and reputation when they pump this out. They think they’ve prompted for thought leadership, but what they’ve got is a potent recipe for egg on face. Now, try this prompt: You are a Financial Times opinion writer. You’re going to write a blog on AI risk in military use. Research the Anthropic vs DoD as a case study. The audience are UK tech innovators who we want to inform about the dangers of misaligned AI. Your style should be similar to Stephen Fry in tone. Aim for 800 words. You didn’t know what to do with that email from your boss, and neither will AI. Both the email and the first prompt did not specify how to approach the task. But if you the boss gave you the second one, they are telling you the role, the task, the context, the audience, the style, and the desired output. These are the core components which you need to consider when you are crafting a prompt. ![](https://storage.ghost.io/c/0b/06/0b06555c-e570-49f3-a2d1-3ff0d72f29c4/content/images/2026/10/eaf7ee44-89b2-463f-853c-8caf1cda78fd_752x293.png) We’ll do a deep dive on context soon, but what you need to know right now is what context is in AI. It is comprised of four things: the instructions you provide (the prompt), the information AI is trained on (knowledge), what it can find online, and any supporting documents you provide (see below for examples). ![](https://storage.ghost.io/c/0b/06/0b06555c-e570-49f3-a2d1-3ff0d72f29c4/content/images/2026/10/712ac6f5-2f44-4942-b514-2ae91db5e799_752x436.png) You can also ask the AI to tell you where it is missing information. This is as simple as putting “Ask me any questions to help you understand this task” at the end of your prompt. ## Modify When you’re interacting with AI, aim for collaboration and co-creation, not delegation and abdication. An easy way to think about this is: “how would I work with another human on this task?” Invite questions, have a back and forth, provide feedback. The role here is of mentor and mentee, not master and slave. This can take place before you get to the prompt as part of a brainstorm – you can even ask AI to prepare a prompt for you post-brainstorm that captures the task. Take this into a new conversation and boom – better outputs. It should also happen after you’ve got the first draft. A really useful quote to have in your head during this process is: “The first draft of everything is shit.” - Ernest Hemingway. Through collaboration, you can transform the piece that you’re working on. If AI’s missed the mark, explain why. Too short, too long, focused on the wrong areas, got the tone wrong: provide all of this as feedback for the next draft. You can also use AI to transform one thing into another. They say a picture is worth a thousand words, so you could upload a picture and say, “Give me a thousand-word blog on this picture.” Likewise, you could take the output from the prompt that I gave you above and say this: > Take this post and transform it so it is written from the point of view of Bender from Futurama, who is trying to cover up his true intentions of using AI to wipe out mankind. Or > Transform this into a satirical op-ed written from the point of view of Pete Hegseth, who is presently drunk and at a frat party, trying to convince people bringing about Skynet is a great idea. Keep going until you’re close to a final draft. ## Verify When you create something with AI and it comes across as generic slop, who is to blame? I won’t leave you in suspense: It’s you. It’s your professional credibility and reputation which are on the line. When reviewing, ask yourself: “If this was published in the FT, would I be happy to have my name against it?” That’s the bar. Nothing in the known universe is infallible, especially humans. That’s why in journalism, we have rigorous editorial processes to catch errors. Check your spelling, double check your sources, triple check your facts. Get someone else to read it before you hit send. Get another AI to review your output for errors and ways to make it tighter. Set your prompt up for maximum specificity so AI doesn’t have to guess at what you actually want. There’s also areas AI can trip on to watch out for: · Training cut off dates. The model will be blind to anything after this date (you can ask yours when its cut-off date is). Check time sensitive facts, and if it looks out of date, ask your AI to do a search for up to date information. · Tokens. AIs don’t read a word or a character: they read in tokens, which are three to four characters. This is why AI struggles with telling you how many Rs there are in Strawberry. Always double check maths. · Being too helpful. AIs are trained using Reinforcement Learning from Human Feedback (RLHF). Unfortunately, most trainers define ‘helpfulness’ as speed over genuinely useful, meaning AI will cut corners. To get around this, always use the thinking mode. It may be slower, but it is far more effective at getting stuff done. This is typically available as a toggle: ![](https://storage.ghost.io/c/0b/06/0b06555c-e570-49f3-a2d1-3ff0d72f29c4/content/images/2026/10/8a04571f-b478-4848-b8e1-f665ab6fbef6_1672x742.png) ## Instructions One of the strongest (yet most underutilised) ways to communicate with AI is personalise it. Known as ‘personas’ or ‘system prompts’, these are instructions which tailor the model to a specific user or use case. For example, both Copilot and ChatGPT use the GPT 5.4 model, but each have different system prompts, giving the user different experiences on each platform. Setting yours will massively alter the way your AI interacts with you. Here’s mine for Claude: ![](https://storage.ghost.io/c/0b/06/0b06555c-e570-49f3-a2d1-3ff0d72f29c4/content/images/2026/10/8542d3b5-e513-430c-aad6-dec4f23b426e_887x245.png) To get yours, use this prompt: > I want you to help me build a set of custom instructions I can paste into your personalisation settings. Interview me first. Ask me questions across these areas (one or two at a time, not all at once): > > 1\. How my brain works — learning style, any neurodivergence, how I process information best 2\. What I actually use AI for day-to-day 3\. My expertise level and professional context 4\. What annoys me most about AI responses 5\. Communication style — formality, length, whether I want challenge or compliance > > Once you’ve got enough, draft a set of custom instructions I can copy and paste. Keep it within the character limit for your personalisation settings. Show me the draft, then let me refine it with you. Have that conversation and get the output. There will be a box in your AI’s settings where you can paste it. In Claude and ChatGPT, you click on your name in the bottom left and click settings, then go to either general (claude) or personalisation (chatgpt). On Gemini, open the hamburger menu (top left) and settings are at the bottom. M365 Copilot is the three dots at the top right. The Specify, Modify, Verify approach – combined with a collaborative approach, thinking mode and custom instructions – will give you much more useful, thoughtful, and robust outputs. Over the next week, try these in combination with the use cases we spoke about in AI Bayesics #1 – you’ll see a massive jump in performance. Next Friday, we’ll examine how to avoid cognitive offloading – the use of AI which rots your brain – and instead use it as a cognitive sparring partner which elevates your thinking to higher levels. [Subscribe](#/portal/signup)