AI Prompting: The Fundamentals

By RJ Militante · April 2026

Most people are prompting AI with the equivalent of a text message. Two sentences. No context. And then they're frustrated when the output is generic.

The core insight behind every technique in this post is this: the quality of your output will never exceed the quality of your input. That's not a motivational statement — it's a direct consequence of how these models work.

These fundamentals apply across every AI tool — Claude, ChatGPT, Gemini, all of it. Get them right and every tool gets better.

How the model actually works

Large language models don't "know" things the way a human expert does. They process massive amounts of text — documents, code, forums, transcripts — break it into tokens, and learn to predict the most statistically likely next token given what came before. It's pattern recognition at scale.

That's why if you type "twinkle, twinkle, little" — it outputs "star." Not because it understands nursery rhymes, but because it has seen that sequence billions of times and learned the association.

This has a practical implication: the more relevant information you put into a prompt, the better it can pattern-match to useful outputs. You can give it the equivalent of two or three books worth of context before asking your question. Most people give it a sentence. That gap is where most prompting fails.

It also explains why generic prompts get generic answers. If your input could have come from anyone, the output will sound like it was written for everyone.

The four-part prompt structure

Every high-quality prompt has four components. You don't always need all four, but the ones you skip tend to show up as gaps in the output.

1. Role

Tell the model who to be.

Act like a world-class marketing strategist focused on conversion for SaaS products.

This narrows the model's entire knowledge domain to a specific lens. Without it, the model tries to serve every possible angle — producing something that's technically accurate and practically useless.

The role does two things: it filters out irrelevant patterns and concentrates the output toward the expertise level you need.

2. Context

Give it everything relevant to the problem. Documents, transcripts, specifications, prior conversations, product details, constraints. The model can hold a lot — use it.

Most prompts skip this almost entirely. The model ends up filling gaps with plausible-sounding guesses. If you have the actual information, provide it. Don't make the model hallucinate what you already know.

3. Command

State exactly what you want. Be explicit.

Make the implicit explicit.

"Help me with my marketing" is a command. "Write three subject lines for a cold email targeting B2B founders in the fintech space — direct, no hype, under 8 words each" is a command. Only one of them has a clear success condition.

4. Format

Specify the output structure before the model starts generating. This is the most commonly skipped step, and it matters more than people think.

Bullet points vs. prose
Table format vs. numbered list
PDF vs. CSV for downstream tool use
A template the model must fill

If you have a specific structure in mind, give it as a template. The model will stay within those parameters. If you don't, it will pick a format based on what it's seen most often for similar prompts — which may or may not match what you need.

Push prompting vs. pull prompting

Most people use AI the hard way. They do 80% of the work — outlining the approach, providing the structure, breaking down the problem — and let AI fill in the last 20%. That's push prompting. You're telling the model how to do the task.

Pull prompting inverts this. You give it the destination and let the model figure out the path.

How to structure a pull prompt

Role + context — who it is, what it knows about your situation
Outcome-based objective — not "write a sales email" but "I need a 5-step email sequence that converts cold leads into booked discovery calls for a B2B SaaS product"
"Ask me all the questions you need to create this" — then specify the format you want the result in
Answer the questions it asks (voice-to-text speeds this up significantly)
Refine: "Ask me more questions to make this more specific"

The model pulls the relevant details from you and builds toward the outcome. You're not writing the structure — you're answering questions.

This is essentially how software engineers have been prompting for complex code generation for years: describe the outcome, not the implementation. The approach is just now becoming mainstream for everything else.

Master prompts

A master prompt is a document that gives the model complete context about a specific role or domain in your life — the equivalent of a manual it can reference for every interaction.

Without one, every conversation starts cold. The model has no idea who you are, what you're building, what constraints matter, or what "good" looks like in your context. Every output is generic until you provide enough specifics to make it otherwise.

With one, you stop re-explaining yourself. The model has your full situation loaded before your first message.

What goes in a master prompt

Use pull prompting to build it:

I want to create a master prompt for my role as [CEO / backend engineer / 
content marketer / etc.]. Ask me all the questions you need to create it.

Answer the questions conversationally. Review the output it generates. Refine anything that doesn't capture the nuance correctly. Save it as a PDF.

When you switch to a different model or tool, upload it. The context transfers. The model goes from stranger to well-briefed immediately.

One practical detail: the more specific the master prompt, the more useful it is. Don't just list your job title. Include your goals, your constraints, your team structure, what decisions you're commonly making, what "good output" looks like for your specific situation, and what context would be obvious to a close collaborator but invisible to a stranger.

System prompts

A master prompt is about you. A system prompt is about a specific workflow.

If you've spent 20 minutes getting an AI to produce exactly the output you needed — the right tone, the right structure, the right level of detail — a system prompt captures that process so you can reproduce it in one click. It's a recipe that reliably produces a specific type of output.

How to build a system prompt

Start with pull prompting:
You're an expert AI engineer. I want to create a system prompt that does X. Ask me all the questions you need to build it.
Answer the questions
Test the output
Iterate: "Refine this, it's missing Y"
Once it's correct, copy the system prompt into a Claude Project, custom GPT, or Gemini Gem

The result is a reusable tool. You give it an input — a topic, a lead's name, a feature spec — and it produces a consistent, high-quality output without any additional prompting. You're building small machines that handle repeatable work.

A public resource worth knowing: a collection of leaked system prompts from production AI tools (Perplexity, Notion, Lovable, and others) was compiled and posted to GitHub. Search "system prompts AI tools GitHub" and you'll find it. Seeing what production-grade system prompts look like is one of the fastest ways to understand what's actually possible.

Pick one tool and go deep

There's a pattern among people who aren't getting results from AI: they're using four different tools interchangeably and haven't built real fluency with any of them. Claude one day, GPT the next, Gemini when someone shares a link.

Learning AI is like learning an instrument. If you split your attention across piano, guitar, and drums simultaneously, you stay mediocre at all three. Pick one, go deep, develop real fluency — then the others become much easier to pick up because you understand the fundamentals.

As of now: Claude leads on code, writing, and long-context reasoning. Gemini is strongest for research and Google Workspace integration. GPT has the most integrations and the largest ecosystem of third-party plugins. Pick based on your actual daily use case, not hype.

The short version

AI is pattern recognition — better input produces better output
Every strong prompt has four components: role, context, command, format
Push prompting makes you do most of the work; pull prompting inverts that
Master prompts give the model persistent context about you and your situation
System prompts capture repeatable workflows so you can reproduce them reliably
Pick one tool, master it before you switch

The gap between people who get consistently good results from AI and people who don't usually isn't model quality or feature access. It's prompt quality.

If you've picked Claude as your tool of choice, the next post goes deeper: Claude Code: A Practical Field Guide covers what changes when you move from chatbot to agent — permissions, context management, CLAUDE.md, plan mode, and more.