🧱 Fundamentals

A set of instructions that reaches the AI before your first message and stays in effect for the whole conversation—its "operating mode," which you don’t see on screen.

Almost everything that frustrates or delights you about an AI starts here. Knowing the note exists is the first step toward no longer suffering through AI and starting to guide it.

Invisible · comes first · stays connected to the entire conversation · is the adjustable part.

Two comparisons: the script defines the character for the same actor; the manual defines how the employee serves customers. The system prompt is both at once.

These are the analogies you’ll carry through the entire course — change the script, follow the manual without reciting it. They anchor everything that comes next.

The script defines the character · the manual defines the rules · same talent · works behind the scenes.

Every conversation has three voices: system (the invisible director), user (you) and assistant (the AI responding).

You see only two, but the third controls everything. Knowing how to distinguish the voices is the basic vocabulary for reading any conversation with AI.

system (director) · user (you) · assistant (the AI) · who wins the conflict.

The two practical reasons for the note: consistency (the AI acts the same way with everyone) and safety (it knows what not to do).

Without a system prompt, AI would be unpredictable and dangerous. Understanding why explains why this text exists in every serious AI system.

Consistency · security · predictable behavior · safeguards.

The same AI, with two different notes, becomes two different things—we saw Grok as a "study buddy" and as a "medical advisor," verbatim.

It’s proof that the ticket is powerful—and adjustable. Watch out for the wrong conclusion: it changes behavior, not knowledge.

Same model · different prompts · it’s the script · it’s not knowledge.

Each AI's "style" starts with the first line of the prompt—we saw Cursor ("You are an AI coding assistant…") and Grok, for real.

That feeling that every AI has its own personality is real — and it’s the choice of whoever wrote the note, not just the model.

Identity on the 1st line · where it operates · stance · is a choice made by whoever wrote it.

The part that says who the AI is, what it is (the model), and even how current its knowledge is (the knowledge cutoff), stated right at the opening.

Without an anchor, AI drifts—it becomes another persona or invents powers. Identity is the first piece and the only one present in 100% of prompts.

Persona · base model · knowledge cutoff · mission scope.

The part that defines the writing style: formal or casual, brief or detailed, with or without lists—the tone calibrated to the product.

It's the element that most changes the "feel" of using AI—and the one that most separates a pleasant chatbot from a robotic one.

Formal vs. casual · prose vs. list · honesty · calibrated tone.

The part that lists what the AI can do in the world: search, read files, run code—each action with its scope and conditions for use.

It’s what turns a chatbot that only talks into an agent that acts. A weak declaration makes the AI promise what it can’t do or ignore what it can.

Action in the world · scope · conditions of use · agent vs. chatbot.

The part that states refusals and limits, placed early to signal precedence: safety comes before usefulness.

A blunt prohibition is easy to work around. A rule with a reason generalizes to situations its author never anticipated — you can see that right here.

Refusal · safety > usefulness · early stance · rule with a reason.

The part that sets precedence when two instructions conflict: central policy > system > user > default.

Without a clear order, conflicting instructions leave behavior to chance—the prompt needs to say who wins.

Precedence · central policy · conflict resolution · user > default.

The way to organize the pieces: each block in a named "drawer" with an XML tag (<tone_and_style>, <tool_calling>), real ones from Cursor.

It’s what makes a long prompt readable and easy to maintain. The drawer’s name is already half the instruction.

Named block · XML drawer · organization · the name already gives instructions.

The first structural pattern: “You are X” + model + knowledge cutoff, seen in 100% of real prompts.

Identity is an anchor, not a biography — recognizing the pattern avoids the antipattern of a buried or bloated persona.

Anchor · "You are…" · 100% universal · short and precise.

The pattern of listing capabilities in a structured way—each with “what it does,” “when to use it,” and “limits.”

A clear list avoids both extremes: offering what doesn't exist and ignoring what does.

Scope · "if the user…" condition · limits · anti-overpromising.

The pattern of wrapping each logical block in a named XML tag—flat (few levels) and modular.

It’s what makes a prompt with thousands of lines maintainable and the diff between versions readable.

Named block · flat · modular · diffable (Markdown works too).

The pattern of pairing an abstract rule with a concrete demonstration (input → output)—like Perplexity’s citation format.

A well-chosen example reinforces behavior better than three paragraphs—the AI learns by imitation.

Demonstration · input → output · imitation · rule + example.

The pattern of choosing meaningful block names—the drawer’s title guides behavior; it doesn’t just organize.

A good name saves lines of rules and sets the stage for the turn in Module 1.6 (the block's name is an instruction).

Meaningful name · guiding label · fewer rules · semantics.

Cases where the XML compartment becomes too much ceremony: short prompts, code (where Markdown headers are enough).

Every pattern has a limit. Knowing when NOT to apply it is what separates those who copy from those who understand.

Pattern limit · too much ceremony · short prompt · Markdown is enough.

The pattern of pairing every rule with the reasoning behind it, instead of a blunt command.

The rationale lets AI generalize to unforeseen cases, instead of obeying mechanically and failing at the edges.

Explicit reasoning · generalization · edge-case robustness · intent over wording.

The pair that sets the desired and prohibited behaviors side by side—we saw Cursor ("NEVER create files… ALWAYS prefer editing…").

The contrast removes the ambiguity in the middle: the AI knows exactly which pole to move toward and which to avoid.

ALWAYS/NEVER · desired/forbidden pair · no middle ground · emphasis.

The explicit instruction about the ideal response length—Grok ("You provide the shortest answer you can…").

Without a budget, the AI "pads the response"; with one, it calibrates density to the channel (terminal, chat, API).

Target length · expand/cut · density by channel · anti-verbosity.

Behavior rules for tools: when to call them, in what order, when to stop — the first look at a topic from Track 2.

Declaring the tool isn't enough — its usage contract is what prevents redundant, out-of-order, or unnecessary calls.

When to call · order · prerequisite · when to stop.

The idea that ALL CAPS, "MUST," and "CRITICAL" are currency: overspending devalues every rule.

When everything is shouting, nothing stands out. Treating emphasis as a scarce resource is what keeps the prompt audible—and anticipates Error 1 in Module 1.5.

Emphasis as currency · inflation · scarce resource · audibility.

The rule for using absolutes (always/never) only when there really are no exceptions — otherwise they clash and lose their force.

Too many absolutes create conflicts that are impossible to obey. Restraint is what keeps every “always” meaningful.

True absolutes · restraint · avoid collisions · preserve strength.

Stuff the prompt with ALL CAPS, "MUST," and "CRITICAL" until the emphasis loses all its value.

It’s the most common beginner mistake. The antidote: treat emphasis as currency and save it for what really matters.

Everything screams · emphasis loses its value · antidote: restraint · audibility.

Adding rule after rule with each problem without ever removing the old ones—the prompt becomes layers of contradictory onion.

It’s how prompts bloat until they become unmanageable. The antidote: consolidate—remove and rewrite, don’t just add.

Layers · internal contradiction · bloat · antidote: consolidate.

Describe the exact mechanics of what to avoid (step by step) instead of the principle—a list that ends up teaching what it meant to prohibit.

It’s the seed of the security lesson in Track 3. The antidote: state the principle, not the recipe.

Mechanics vs. principle · dangerous symmetry · antidote: the why · high-level.

Put what should be current context (date, user, task) into the stable prompt — mixing what doesn't change with what changes each turn.

Confusing the layers leads to fragile prompts and wrong conclusions when reading a leak. The antidote: separate fixed from dynamic.

Stable vs. dynamic · current context · harness · antidote: separate them.

Specify every tiny step out of distrust, without letting the AI use its own judgment — instructions that suffocate instead of guide.

Micromanaging breaks generalization. The antidote: trust the AI’s judgment by giving it principles instead of step-by-step recipes.

Micromanagement · stifled judgment · loss of generalization · antidote: trust.

The wrap-up that brings the 5 errors and their antidotes together in a single table, with "emphasis is currency" as the guiding thread.

Seeing all five together turns a list of cautions into a mental checklist you can apply to any prompt.

Emphasis is currency · 5 mistakes + 5 antidotes · mental checklist · synthesis.

The field’s change in direction: from prompts packed with rigid rules to prompts built on generalizable principles.

Understanding the trend puts you ahead — you start writing in the style the labs are moving toward, not what’s already outdated.

Trend · rule → principle · field maturity · labs’ direction.

A real diff between two versions of a lab prompt—the annotated comparison that shows the shift happening in practice.

Reading changes through diffs is the core skill of Track 3 — here you get your first look at it in a concrete, narrated example.

Version diff · real case · annotated comparison · read the evolution.

The case’s paradox: the prompt gained 56 lines but lost 7 rule blocks—it grew in explanation and shrank in mechanics.

It challenges the intuition that "bigger = better." The right direction is to consolidate rules into principles, even if the text gets longer.

Evolve ≠ grow · +56 / −7 · consolidation · quality > size.

In the real case, copyright_compliance became core_copyright_principle — renaming alone changes the AI’s stance.

From "follow a rule" to "honor a principle": the name carries the intent. Small change, big effect.

Renaming instructs · compliance → principle · stance · intention in the name.

The view that every block removed between versions reveals a decision—what was removed matters as much as what was added.

Beginners only look at what was added. Learning to read deletions is half the skill of interpreting a diff.

Removal = information · decision revealed · read what was removed · double signal.

The practical takeaway: when writing or adjusting a prompt, always prefer a principle that generalizes over a rule that covers only one case.

It's the compass that closes Track 1 and opens Track 2—you'll be ready to get hands-on with the right mindset.

Principle > rule · generalization · practical compass · bridge to Track 2.