DAN MARTELL · EXTRACTED

You’re not behind (yet): How to learn AI in 18 minutes

The four-part prompt structure, the pull method, and why most people are doing 80% of AI's job for it.

"The quality of your output will never exceed the quality of the input to your prompt." — Host
Host (unnamed), on the episode

This is a solo breakdown from a founder who has tested thousands of AI prompts across every major platform, built custom AI tools for his own companies, and worked with hundreds of founders on implementation. The pop framing around AI is that the tool is the problem: people switch between ChatGPT, Claude, and Gemini hunting for better outputs. The actual problem is almost always the prompt. This episode surfaces the operational framework underneath: how to structure a prompt so the model has enough context to produce something useful, how to flip from push prompting to pull prompting so AI does the heavy lifting instead of finishing your sentences, and how to codify repeatable outputs into system prompts that work without re-explaining anything. These are the plays that separate the top 1% of AI users from everyone else.

Tactic 01

The Four-Part Prompt Structure

Most people give AI a text message and expect magic. The host describes working with a founder who spent three weeks trying to get his marketing copy right, iterating back and forth and never landing the headline. The host sat down, applied a four-part structure, hit enter, and within 12 seconds the model returned the complete plan, structure, and strategy. The founder implemented it in under 15 minutes and started converting traffic into customers. The structure has four components. First, the role: tell the model who to be. Not just 'marketing expert' but something specific like 'a world-class marketing strategist that focuses on conversion for software.' This narrows the model's entire knowledge base to the relevant slice. Second, context: give it all the background material you have. Transcripts, documents, specs, anything. The more it has, the better it can guide its output. Third, the command: be explicit. Make the implicit explicit. Tell the model exactly what you want it to do. Fourth, format: specify how you want the output returned. A bulleted list, a table, a CSV, a PDF. If you want it to follow a template, paste the template and tell it not to go outside those parameters. The host notes you can go as granular as providing a literal format template for the model to fill in.

The play

Before your next AI request, write out all four components before hitting enter: role, context, command, and format. For the role, be specific about the discipline and the outcome focus, not just the job title. For context, paste in every relevant document, transcript, or spec you have access to. For the command, make every implicit assumption explicit. For the format, either name the output type or paste a template. Run this once on a task you have been getting weak results on and compare the output to what you were getting before.

Tactic 02

Pull Prompting: Give the Destination, Let AI Drive

The host draws a sharp line between how most people use AI and how it should be used. Push prompting is doing 80% of the work yourself and letting the model finish the last 20%. You are telling AI how to do the task. Pull prompting inverts this: you give the destination and let the model figure out the best route. The analogy he uses is turn-by-turn directions from a friend versus GPS. Push is the friend telling you to turn left at the church and right at the red mailbox. Pull is typing in the address and letting the system navigate. The pull method runs in three steps. First, give the model the role and context as normal. Second, state the outcome-based objective clearly, something like 'I need a sequence that converts cold leads into booked calls.' Third, add this exact instruction: 'Ask me all the questions you need to create this for me, then give it back to me in this format.' The model returns a list of questions. You answer them, and then it builds the output from your answers rather than from its own assumptions. The host recommends using voice-to-text to answer the questions since the model may ask seven at once and typing them all out is slow. He notes this is how software programmers have been prompting for a while under the name spec-driven development, and it is now accessible to anyone.

The play

On your next complex AI task, do not describe how to do it. Instead, state the outcome you need in one sentence, then add: 'Ask me all the questions you need to create this for me, then return the result as a formatted document.' Use voice-to-text to answer the questions the model returns. Once you have the draft, tell the model to ask more questions if it needs to refine it further. Repeat until the output matches what you actually need. This shifts the cognitive load from you to the model and produces outputs that are specific to your situation rather than generic.

Tactic 03

The Master Prompt: Introduce Yourself to AI Once

Without a master prompt, every conversation with AI starts from zero. The model has no idea who you are, what you sell, how your life is structured, or what matters to you. The result is generic auto-completion that could apply to anyone. The host's fix is a master prompt: a document that functions like a manual about a specific role in your life. If there were a manual called Dan CEO, it would contain everything about the business, the team, what is sold, how decisions get made. If there were one called Dan Father, it would contain the people in that life and how things are structured there. He requires every team member in his companies to have a master prompt and to show it to him. He cannot accept that someone is chatting with AI all day and getting useful outputs without the model having this context. Building one uses the pull method: tell the model you want to create a master prompt for your role as whatever your role is, then ask it to give you all the questions it needs to create it. Answer the questions by voice. Review the output and refine anything that does not feel accurate by talking to the model directly. Save it as a PDF named 'master prompt' plus the role name. When you switch tools or open a new conversation, upload it so the model has full context from the first message.

The play

Open your AI tool of choice and type: 'I want to create a master prompt for my role as [your role]. Ask me all the questions you need to create it.' Answer every question using voice-to-text. Once you have the document, read through it and tell the model to update anything that is missing or wrong. Save it as a PDF. From now on, upload it at the start of any new conversation or when switching to a different tool. You only have to build it once per role, and it converts every future interaction from generic to specific.

Tactic 04

Structure Pull Prompts With Outcome-Based Objectives

Most people prompt AI the same way a friend gives you directions. Turn left at the house, right at the church, look for the red mailbox. You do all the thinking, and AI just fills in the last step. That is push prompting, and it means you are doing 80% of the work while AI finishes the easy 20%. Pull prompting flips the ratio. You give AI the destination and let it figure out how to get there. The structure has five steps. First, give the role and context as you normally would. Second, state an outcome-based objective, something specific enough to function as a destination. The example given in the video is: "I need a sequence that converts cold leads into booked calls." Third, add this exact instruction: "Ask me all the questions you need to create this for me, then give it back to me in" whatever format you want, a table, a PDF, a numbered list. Fourth, answer every question AI comes back with. A practical shortcut here is voice mode: tap the microphone and answer conversationally rather than typing out responses to seven questions in a row. Fifth, once you have the output, you can optionally loop back and say, "Ask me more questions to make this more relevant," and keep refining until it lands. The reason this works is that it matches how experienced software developers have been using AI for some time. As the speaker puts it, coders are "coding through outcomes," a practice called spec-driven development. Pull prompting brings that same approach to anyone, regardless of technical background. You are not telling AI how to do the task, you are telling it what you need and letting it surface the right questions to get there.

The play

Write a pull prompt in three parts. Open with a role and brief context. Then state your outcome-based objective in one specific sentence, something concrete enough that you could call it a destination. Close with: "Ask me all the questions you need to create this for me, then return it as [format]." When AI responds with questions, use voice-to-text to answer each one conversationally. Once you have the first output, run one refinement pass: "Ask me more questions to make this more relevant." The whole sequence, from first prompt to finished output, should take under 15 minutes on most tasks.

Tactic 05

Pick One Daily Task For AI Automation

Most people finish a video like this, feel motivated for about six minutes, and then go back to doing everything the way they always did. The reason is not laziness. It is that the gap between understanding something and actually using it never gets bridged by a single small action. The host is direct about the fix: pick one daily task and have AI do it. Not ten tasks. Not a system overhaul. One task. The example he gives is as unglamorous as it gets: moving information from one system to another. Not writing a business plan, not generating a marketing strategy. Just one repetitive thing you already do every day, handed off to AI for ten minutes. The point is that the habit of reaching for AI before doing the task yourself is what compounds. You are not optimizing the task. You are building the reflex. As he puts it, "10 minutes of action beats 10 hours of tutorials."

The play

Today, identify one task you do on repeat, something you did this week that felt mechanical. Moving data, drafting a routine reply, summarizing a document. Give it to your chosen AI tool with a proper prompt using the role, context, command, and format structure. Set a ten-minute timer and let it run. Do this for the same task every day for one week. The goal is not a perfect output on day one. The goal is that by day seven, reaching for AI before doing the task yourself feels automatic.

Tactic 06

Deploy System Prompts In Custom GPT Or Claude Projects

Most people treat AI like a conversation they have to restart every time. They spend 17 minutes getting to a great output, then lose the thread the next day and start over. System prompts solve this. The idea is simple: once you have an output you love, you reverse-engineer the instructions that produced it into a reusable recipe. As the host puts it, "You code system prompts with words." The role, the context, the format, the constraints, all of it gets captured in one document that runs the same way every time you trigger it. Once the system prompt is dialed in, you copy it into a custom GPT, a Claude project, or a Gemini gem. From that point, the complexity disappears for everyone using it. You can hand someone a link and say, as he describes, "Just give it the topic and it'll do something for us." The system carries the intelligence. The user just feeds it the variable. This is how you stop being the bottleneck on every AI-assisted task your team runs. The build process follows the same pull structure. You tell the AI, "You're an expert AI engineer. I want you to create a system prompt that does X. Ask me all the questions you need to create it." You answer conversationally using voice-to-text. You test the output, refine anything off, and then lock it in. If you want to see what world-class system prompts look like in practice, the host flags a real shortcut: a leak collected the full system prompts from tools like Perplexity, Notion, and Lovable. Search "system prompts and models of AI tools" on GitHub and you can study exactly how the best-built AI products instruct themselves.

The play

Pick one repeatable task you use AI for at least three times a week. Open your AI tool and say: "You are an expert AI engineer. I want you to create a system prompt that does [specific task]. Ask me all the questions you need to create it." Answer every question using voice-to-text so you don't slow down. Test the system prompt it returns, refine anything that feels off, then copy the final version into a custom GPT, Claude project, or Gemini gem. Share the link with anyone on your team who runs the same task. They give it the input, the system does the rest, and you never have to re-explain the setup again.

Tactic 07

Pick One AI Tool And Master It Deeply

Most people bounce between Claude, ChatGPT, and Gemini, convinced they are covering their bases. The host is direct about what he actually sees when those people show him their work: "Show me your prompts. And it's Crayola. It's simple. They haven't mastered the tool." Switching platforms is not progress. It is avoidance dressed up as curiosity. He frames it the same way he would frame learning an instrument. If you split your practice time between piano, guitar, and drums, you end up mediocre at all three. Pick one and go deep, and the others become accessible later because you understand the underlying music. The same logic holds for AI. Once you understand how one platform thinks, how it responds to context, how it structures outputs, the others become far easier to pick up. The host names three platforms worth committing to: Claude for writing, creative work, deep thinking, and code; Gemini for research and anyone living inside Google tools; ChatGPT for integrations and breadth. Each has a lane. Pick the one that matches your current work, then stay there.

The play

Choose one platform based on your primary use case and commit to it for at least the next 90 days. If you write, think, or code, start with Claude. If you research or live in Google tools, start with Gemini. If you need integrations or a proven general tool, start with ChatGPT. Do not open the others during this period. Every hour you spend switching is an hour you are not building the fluency that turns a tool into a weapon.

Tactic 08

Load AI With Two To Three Books Worth Of Information

Most people treat AI like a search bar. They type a sentence, maybe two, and then wonder why the output feels generic. The host is direct about this: "Did you know you can give it like two, three books worth of information about your question before it even gives you an answer? Most people are giving it a text message. It's giving it nothing. And then you're expecting it to do magic." This isn't a metaphor. It's a description of how the model actually works. AI builds context from what you give it. The more signal you load in before the question, the more precisely it can pattern-match toward a useful answer. Marketing documents, call transcripts, product specs, research notes: all of it can go in, and all of it improves the output. The host uses the analogy of someone telling you every detail of a story before asking you a specific question. You get it right because you have the full picture. AI works the same way. The practical implication is that the gap between a mediocre output and a great one is almost always in the input, not the model. When the host sat down with the founder who had spent three weeks stuck on his marketing messaging, he did not use a smarter tool or a more creative approach. He loaded the prompt with rich context, used the four-part structure, and hit enter. The plan came back in 12 seconds. "It took less than 15 minutes to implement it on his website and all of a sudden he's converting that traffic into customers." The model had not changed. The input had.

The play

Before you write your next prompt, spend five minutes gathering every relevant document, transcript, or background note you have on the topic. Paste it all in before your actual question. Think of your prompt as a briefing, not a request. If you are asking AI to help with sales copy, load in past customer calls, your product one-pager, and any competitor messaging you have on hand. The more specific context you give it, the less it has to guess, and the less you have to revise afterward.

Your action plan

the plays, one checklist.

01
The Four-Part Prompt Structure
Structure every prompt with four parts: role (who should it be), context (all background material), command (exact explicit instruction), and format (how you want the output). Paste a template if you have one. The model cannot exceed the quality of what you give it.
02
Pull Prompting: Give the Destination, Let AI Drive
State your desired outcome, then tell the model to ask you every question it needs before producing anything. Answer the questions by voice, let it build the output from your answers. You do 20% of the work; it does 80%.
03
The Master Prompt: Introduce Yourself to AI Once
Use pull prompting to build a master prompt document for each of your main roles, save it as a PDF, and upload it at the start of every AI conversation. You build it once; every future output becomes specific to your actual situation instead of generic.
04
Structure Pull Prompts With Outcome-Based Objectives
Write a pull prompt with three parts: role plus context, a single specific outcome-based objective, and the instruction 'Ask me all the questions you need to create this.' Answer AI's questions by voice if you can. Run one refinement pass at the end.
05
Pick One Daily Task For AI Automation
Name one mechanical task you did this week and give it to your chosen AI tool today using the four-part prompt structure. Do it for ten minutes. Repeat daily for one week to build the reflex, not just the skill.
06
Deploy System Prompts In Custom GPT Or Claude Projects
Pick a repeatable AI task, build a system prompt using the pull method, then deploy it inside a custom GPT, Claude project, or Gemini gem. Hand the link to your team. They feed it the input, the system handles the rest, and you never rebuild the setup again.
07
Pick One AI Tool And Master It Deeply
Pick one platform that matches your work, Claude for writing and code, Gemini for research, ChatGPT for integrations, and commit to it for 90 days. No switching. Fluency compounds only when you go deep in one place before you go wide.
08
Load AI With Two To Three Books Worth Of Information
Before writing your prompt, paste in every relevant document, transcript, or background note you have on the topic. Treat the prompt as a briefing, not a request. The more specific context you load, the less AI has to guess and the less you have to fix afterward.