A friend came to talk to me about AI recently. To save them the embarrassment, let’s call them Sam.
Sam is sharp. Runs a business that does millions in revenue, manages life with the same discipline they run their company. The kind of operator you’d back to figure most things out on their own. So when they called me last week and said “I just don’t get AI,” I was surprised.
They’d tried it. A Claude and ChatGPT subscription with a few months of exploration and experimentation. The results were what they called “Okay, I guess.” Drafts that took longer to fix than they would have taken to write from scratch. Generic copy filled with AI slop. Summaries that missed the point. Emails that sounded like a robot pretending to be a person. They couldn’t see what all the fuss was about. They were ready to cancel the subscription.
Sam isn’t alone. Sam is the majority. Every survey on AI adoption comes back the same way. The gap between AI’s promise and the average user’s actual experience is enormous. People hear about a 10x productivity boost. They get 1.2x, maybe. Then they assume the promise was just hype all along.
The promise wasn’t hype. They just haven’t been shown how to use it properly.
This is the guide I wish someone had written for me when I started.
Show, don’t tell
I’m not going to give you a list of high level tips and leave it at that. Most articles do that. You read the list, nod along, then go back to doing exactly what you always did.
So I’m going to do something different. I’m going to write this piece in front of you. Every prompt I’m sending to Claude. Every move I’m running on the output. Every correction I’m making. You watch the process, judge the result, and decide if the method has merit.
If you want the companion piece on which tools I use day to day, I covered that in You’re Probably Not Getting the Most Out of AI. Read them together. That article is focused on what tools to use. This one is focused on how to think.
Before we go in, here’s an outline of what we’re about to do.
I run four moves on every AI task that matters. The difference between running them and not running them is the difference between AI you can ship and AI you delete.
These four moves are as follows (see attached image):
The order matters because each move makes the next one work harder. Better context produces sharper briefs. Sharper briefs surface fewer gaps in the reverse interview. Fewer gaps mean the three passes on the first draft gives a more polished output. These techniques are designed to be used together. Skip any one of them and the rest of the work has to compensate.
We’ll walk through each move as we write this article together. By the end, you’ll have watched all four in action and produced this piece with me.
Let’s begin.
1. Context is the lever
The first thing nobody tells you is that without context AI is cut off at the knees. You ask a generic question, you get a generic answer, you blame the AI. However, the fault often sits with the user and the prompt.
Spend the time on the front end and it will make for drastically less editing on the back end.
PRODUCTIVITY HACK: One tool worth mentioning before we go further.
Wispr Flow is a speech-to-text tool that lets me dictate context at the speed I think. Typing a 1,000-word brief takes me 30 minutes. Speaking it takes 5. The reason context discipline is so rare isn’t that operators don’t believe in it, it’s that typing full context takes too long. Fix the input bottleneck and the rest gets easier. For all of the below activities I recommend trying them with Wispr Flow to remove the overhead of having to type and the bottleneck of speed it creates.
CREATE REUSABLE CONTEXT: A critical investment in your AI set up.
Before I write a single word with Claude, I’ve already loaded it with context. I have a file called AboutMe.MD. It’s the operating manual for AI to work with me and once set up, I can reuse it across all my workflows.
It says (in a much more detailed manner) I’m the CEO of Visory and write a Substack called AI from the Inside.
It describes my voice: direct, operator-first, no em-dashes, Australian English.
It lists the words I never use and the cadence I prefer.
It describes my audience.
It contains the running notes from past articles, past prompts that worked, past mistakes that didn’t.
Every serious conversation I have with AI starts by loading that file. The model immediately has a 90% better picture of who’s asking and what they actually want.
Building this file is the highest-leverage 20 minutes you’ll spend with AI this year. Most people skip it because it feels like overhead. It isn’t overhead, it’s the foundations for success.
Here’s how to build one. Skip ahead if you already have yours set up.
Build your AboutMe.md in 20 minutes
Think of it as the brief you’d write for a brilliant new employee on day one. A working document that tells the model who you are, what you care about, and what good looks like in your world.
The .md just means markdown. A plain text format that AI reads more reliably than Word docs or PDFs. Any notes app or text editor will do.
Keep it short. Three to five hundred words. One screen’s worth. Anthropic’s own guidance on context files is to keep them under 200 lines and resist the urge to write more. Studies have found that longer context files actually reduce output quality and increase cost by over 20%. Small and sharp wins.
Five sections that are critical to your AboutMe.MD file:
Who you are and who you serve. A few sentences on your role. A few sentences on your audience. Without this, the model defaults to generic professional.
What you’re working on right now. Two or three north-star priorities. Not your Tuesday to-do list. The big-picture things that frame how you make decisions.
Voice rules, positive. A list of specific rules on how you like to write. “Short sentences. Plain words. Lead with the conclusion.” Vague is worthless, specific is effective.
How to work with me. Should it ask clarifying questions before drafting, or attempt and flag uncertainty? Default response length? When unsure, what do you want it to do?
What you reject. Counterintuitively, this is the highest-leverage section. The words you never use. The phrases that scream AI-generated. The formatting habits you hate. Be specific: “No em-dashes. No bullet lists for narrative. Never start with ‘In today’s world.’ Avoid ‘delve,’ ‘leverage,’ ‘navigate.’” This section kills generic output more than any other.
IMPORTANT: Don’t literally write it from scratch. Treat AI as a design partner for creating this file. The prompt below does most of the lifting. It audits the context the model already has on you (Project files, Custom Instructions, prior conversations), drafts what it can fill with confidence, and asks targeted questions only about the gaps.
CHEATSHEET: To build your own AboutMe.MD file, copy and paste this into Claude or ChatGPT:
I want you to help me build an AboutMe.md file that I'll load into every future conversation. The purpose of the file is to give you enough context about me, my work, and my voice that your output is dramatically more useful from the first response.
## Step 1: Audit
Audit what you already know about me from these sources, in order:
1. Anything I've loaded into prior conversations (uploaded files, pasted content, attached documents)
2. Any persistent context you have access to (Project Instructions, Custom Instructions, memory from prior conversations)
3. Anything I've written in our conversations so far
State explicitly which of three states applies:
- "Rich context. I can draft most of the file."
- "Partial context. I can draft some sections."
- "No context. Starting from scratch."
If state three, skip Step 2 and Step 3. Go directly to Step 4.
## Step 2: Draft
Draft my AboutMe.md using the structure below. For each section:
- High-confidence fill from the audit: fill it.
- Low-confidence inference: write it and mark [CONFIRMATION, please verify or correct].
- No information: write [GAP, I'll ask you about this].
[GAP] and [CONFIRMATION] markers are placeholders inside the draft. They are not questions. They do not satisfy Step 4.
Structure:
1. **Who I am and who I serve.** My role, my business or organisation, the audience I write or decide for.
2. **What I'm working on right now.** Two or three north-star priorities that frame how I make decisions.
3. **Voice rules, positive.** Three to five specific rules about how I want you to write. Sentence length, word choice, structure, tone.
4. **How to work with me.** Should you ask clarifying questions before drafting, or attempt and flag uncertainty? Default response length? When unsure, what should you do?
5. **What I reject.** Words, phrases, formatting habits, and structural moves I never want to see. AI tells, generic phrasings, em-dashes, anything that signals "written by a model."
Draft constraints:
- Markdown, clear section headers, bullets where appropriate
- 300 to 500 words. Hard cap at 500.
- No filler. Every line should be usable for a real decision about my output.
Deliver the draft as a single code block I can copy.
## Step 3: Sources
List which sources filled which sections. Be explicit. Example: "Section 1 from loaded project file. Section 3 inferred from writing sample. Section 4: no information."
## Step 4: First question
Ask me exactly one question. Then stop and wait for my answer.
If you drafted the file, ask about a [GAP] or [CONFIRMATION] item.
If you have no context, ask the first question for Section 1.
Questioning rules for the whole interview:
- One question per turn. Wait for my answer before the next.
- If my answer is vague ("I want clear writing," "I work with executives"), reject it and demand a specific example.
- Me: "I want my writing to be clear and direct."
- You: "Clear how? Give me one sentence you'd love to see and one you'd reject."
## Required output order
Your response is not complete until all of the following exist, in order:
1. The audit statement (one of three states)
2. The draft in a single code block (skip only if state three)
3. The sources breakdown (skip only if state three)
4. Exactly one explicit question directed at me, followed by a full stop on your turn
If you produce 1, 2, and 3 without 4, you have failed the task. Listing [GAP] markers inside the draft does not count as asking a question. Do not end your response without an explicit question.One thing worth noticing in this prompt: it forces the AI to declare which sources it used for each section, and to flag low-confidence guesses as questions rather than assertions. That’s a move worth running on any important AI output, not just this exercise.
Where to load the completed AboutMe.MD file? Simplest implementation for beginners: In Claude, create a Project and paste the file into Project Instructions. Every conversation inside the Project starts with the file loaded. In ChatGPT, drop it into Settings, Personalisation, Custom Instructions. For one-off chats on any platform, paste the file at the top and say: “Treat the contents of this file as authoritative context for our conversation. Read it carefully before responding.”
Maintain it like a living document. When the model gets something wrong twice, add a rule that would have prevented it. Every couple of months, skim it and cut anything stale. The leanest file that covers what you actually need is always the best one.
2. Tell it exactly what you want
The second move is obvious in theory and rare in practice. Tell the AI exactly what you want.
Bad:
“Write me a Substack post covering best practices for outsourcing your work to AI.”
Better:
“Write a 1,800-word Substack post for AI from the Inside on best practices for outsourcing your work to AI. Audience: operators and busy founders who haven’t yet figured out how to get real value from AI tools.”
Best:
“Write a 1,800-word Substack post for AI from the Inside. Topic: how to outsource your work to AI without outsourcing your thinking. Hook: a story about a friend called Sam who tried AI and was underwhelmed. Body: the method I use to get real value, walked through as a worked example of writing this very piece. Create a 4 step framework for this. Close: the gap between AI users and prepared minds, and why AI favours the prepared mind. Voice: direct, operator-first, no em-dashes, Australian English. Length: 1,800 to 2,200 words. The reader should walk away with a method they can run tomorrow.”
The third version is the kind of brief AI will respond to. It tells the AI the genre, the audience, the structure, the tone, the length, and the desired outcome. Notice how much of it has nothing to do with the topic itself. That’s the point. Most of your specificity should be about the shape of what you want, not the subject.
PRO TIP: Treat the model like a high-performing contractor you just hired. You wouldn’t tell a freelancer “write something about churn” and expect to get back what was in your head. Stop doing it to the model.
The same gradient works for any task. Here’s the move applied to a different brief, an internal memo on customer churn:
Bad:
“Write something about why our customer churn is up.”
Best:
“Write a 400-word internal memo to my exec team explaining why churn is up this quarter. Three paragraphs. Lead with the conclusion. Use the data I’m pasting below to inform your analysis and approach. Sound like a CEO, not a consultant. End with the three things I am asking each of them to do this week.”
The bad version produces something you delete. The good version produces something close to final on the first pass. Same topic. Same model. Completely different outputs, because one version told the model what it actually needed.
BONUS: Download the PDF cheat sheet on the 4-Move Framework to reference whilst working.
Want to support me? A like and comment for reach is much appreciated :)
3. Run a reverse interview
Here’s a technique most people never invoke. Before you ask AI to produce anything, ask it to interview you.
The prompt is simple:
“Before you draft this, what do you need to know that I haven’t told you? Ask me 5 to 10 questions that would help you produce a much better first draft.”
Claude comes back with questions like:
What’s the single insight you want the reader to leave with?
Are there specific moments from your own experience you want included?
What’s the relationship between this piece and your previous work on the topic?
What would success look like for this article: subscribers, comments, shares?
Are there things you specifically don’t want to say or angles you want to avoid?
Then, use Wispr Flow to respond to each of these questions by voice-to-text, saving you hours of time typing and editing.
Now I’m not just briefing the AI. The AI is briefing me. It’s surfacing my own gaps before I’ve written a word. Half the time, the questions force me to think through things I’d taken for granted. The brief gets sharper and therefore your first draft gets stronger.
This is the single best move I’ve added to my workflow in the last year. It takes five minutes and saves hours of rewriting downstream.
4. Run these three moves on every draft
Once the first draft lands, the real work begins. The first draft is never the final draft, it’s simply a starting position.
I run three moves on every output. Each one takes a few minutes. Each one materially improves what comes back.
Move 1: The Skeptic Pass
Paste the draft back into the conversation and write:
“Read this as a sceptical version of [the target reader]. What’s weak? Where would they push back? Give me five honest critiques as if you were that reader trying to reject this draft.”
You’ve just turned the AI into the harshest possible reviewer of its own work. The critiques come back specific and brutal. “The opening assumes the reader cares about Sam; you haven’t earned that yet.” “The middle section repeats the same idea three different ways.” “The close doesn’t deliver on the promise of the hook.”
Then I tell Claude to fix what Claude just identified. The second draft is dramatically better than the first.
The Skeptic Pass is the move that changed everything for me. For a long time, I used AI and accepted what came back. The output was technically correct, structurally sensible, and intellectually empty. In other words: AI slop. I got sick of it and started telling the models so. I’d paste the draft back and say “This is slop. Tell me why.” The models, it turns out, can recognise their own slop perfectly well. They just don’t volunteer the assessment, you have to ask.
The model is much better at self-critique than self-generation, once you give it a role to inhabit.
Move 2: The Blind Spot Review
This one is even simpler:
“What did I forget to tell you in my brief that would have made this draft stronger?”
This is where you find out you left out something critical. A constraint. An audience nuance. A piece of background that changes the whole shape of the answer.
The answer is almost always something concrete. “You didn’t tell me your target reader’s level of AI familiarity, which would have changed how much background I included.” “You didn’t mention whether you wanted a call to action at the end.” “You didn’t give me an example of a previous piece in this voice that I could match.”
Add the missing context. Restart the draft. The next version is better again.
The first time you run the Blind Spot Review you’ll feel slightly stupid. By the third time it’ll be one of the most useful questions in your workflow.
Move 3: The Benchmark Test
Paste in a piece of your own work that actually landed. A post that performed. An email that got the response you wanted. A memo that cut-through and moved a decision forward.
“Compare your draft to this. Why is yours worse? Rewrite to match the structure, confidence, and voice.”
This is the move most writers skip and the move that compounds fastest. You’re not just teaching AI what good looks like in the abstract. You’re teaching it what good looks like for you. The drafts start landing in your voice without you having to edit them sentence by sentence.
You don’t need to articulate what made the gold piece good. The model can see it.
Together, these three moves turn an average first draft into something publishable in about ten minutes of work. Compared to writing it from scratch, that’s maybe a quarter of the time. And the bigger gain isn’t speed. It’s that the final output is better than what you would have produced alone, because the Skeptic Pass and the Benchmark Test have you holding a higher standard than your tired Tuesday-afternoon brain would have set.
If you take one thing from this piece, take this:
You can outsource your work to AI. You can’t outsource your understanding.
The work used to be the typing, the structuring, the summarising, the drafting, the comparing, the formatting. AI is genuinely better at all of it than you are. Hand it over.
The understanding is yours. The judgement of what good looks like. The knowledge of the room you’re writing into. The instinct for which 20% of the draft is wrong and why. The standard you’re holding the output to. Overall, we call this taste. None of that can be delegated. The moment you try to delegate it, you get Sam’s original result. Generic output. Average quality. Disappointment.
AI favours the prepared mind. The reader who skipped this article and went straight to Claude or ChatGPT will get worse output tomorrow than the reader who spent ten minutes building an AboutMe.md file and writing a real context brief. Not slightly worse, but dramatically worse. The compounding gap over time on this becomes enormous.
Everything above is just a way of pushing more of your thinking into the model before it picks up the pen.
Where this leaves Sam
I sent Sam a draft of this article before I published it. Told them to read it, then go back to Claude and ChatGPT and try the moves. They came back a week later. The message was short.
“Mate, much better. The quality has improved out of sight.”
The promise of AI is real. The reason most people aren’t seeing it is the quality of their context and brief, not the model. Fix the context and the output will fix itself.
The gap that matters in the next two years isn’t between people who use AI and people who don’t. It’s between prepared minds and unprepared ones. Changes will keep happening and what counts is your ability to shift with them and learn something new, every day. Which side you’re on is determined by how much thinking you bring to the conversation.
BONUS: Download the PDF cheat sheet on the 4-Move Framework to reference whilst working.
If this resonated, share it with someone who is navigating the same shift.
Want to support me? A like and comment for reach is much appreciated :)
Best, Thomas
thomaspaule.com.au · visory.com.au · LinkedIn
Thomas the last steps on getting AI to check its own work are great. Have you tried telling AI to do all of those checks in one pass or do you have to let it run one at a time?