Prompt

A prompt is the input text that a person or a system gives a large language model (LLM) so it produces an output: an answer, a summary, an email, a classification or an action. It is the basic unit of communication with generative AI, and it works as a natural language instruction that the model interprets to decide what to generate. Unlike programming with code, here the "program" is a sentence or a set of instructions written in Spanish or English.

The quality of a prompt largely determines the quality of the response. A vague prompt ("tell me about sales") leaves everything to chance; a precise prompt (role, task, context, format and constraints) narrows the space of possible answers and produces something useful. That is why, in production systems, the prompt is not improvised in each conversation: it is designed, tested and versioned, a practice known as prompt engineering.

In an AI agent, the prompt stops being a standalone message and becomes a structured layer: there is a system prompt that sets the behavior, the available tools and the rules, and a user prompt that brings the specific request. That orchestration of instructions is part of what goes into designing AI agents that automate sales, support or document tasks on Salesforce.

When someone types "write a follow-up email for a customer who did not reply to the quote", that text is a prompt. The model receives it as input, internally converts it into tokens and, based on patterns learned during training, predicts the most likely sequence of text as a response. The prompt is, then, the main lever the user has to steer a system that, at its core, does not "understand" in the human sense but rather completes sequences statistically. This explains why the same model can give brilliant or useless answers depending on how it is asked.

Anatomy of a good prompt

An effective prompt is not a mysterious formula or a long sentence: it is a structure. In professional settings, it pays to include several explicit elements:

• Role or persona: who the model should "be" (for example, "act as a collections analyst").
• Task: the specific, single action being requested (summarize, classify, write, compare).
• Context: the relevant data (the customer, the history, the balance, the region).
• Output format: whether you want a table, a list, JSON or a paragraph.
• Constraints: tone, length, language, what to avoid and what not to make up.
• Examples: one or more samples of the expected result, a technique known as few-shot prompting.

The more explicit these elements are, the less room there is for ambiguity. A prompt that says "respond in Rioplatense Spanish, 80 words max, without making up data that is not in the context" produces far more predictable results than one that only says "respond well".

Why it matters for the business

In a company, the prompt is what separates a toy experiment from a reliable system. Picture a consumer goods team in Argentina that wants to classify thousands of WhatsApp inquiries by intent (complaint, order, price inquiry). The model is the same for everyone, but the prompt that defines the categories, the examples and the edge-case rules is what makes the classification accurate 95% of the time or barely 60%. The prompt is operational intellectual property: it captures the company's knowledge of how it wants the AI to behave. That is why it is versioned, documented and its performance measured like any other software asset.

A concrete example (Latin America)

A wholesale distributor wants to automate the answer to "when will my order arrive?". A naive first prompt would be: "Tell the customer when their order will arrive". The result is useless because the model does not have the data. The production prompt, by contrast, gives the agent a tool to check the status in the system (see tool calling) and instructions such as: "If the order has shipped, share the estimated arrival date in a friendly tone; if it is delayed, apologize and offer a contact channel; never promise dates that are not in the data". That leap, from asking for a piece of text to designing the agent's behavior, is the essence of professional prompting.

Common mistakes

• Overloading the prompt: packing five different tasks into a single instruction degrades quality. It is better to have one task per prompt or to chain several.
• Assuming context the model does not have: the LLM does not know your database unless you pass it in the prompt or through data retrieval (RAG).
• Not setting anti-fabrication rules: without explicit instructions, the model can produce a hallucination, that is, confidently make up data.
• Ignoring security: a poorly hardened prompt is vulnerable to prompt injection, where malicious text tries to hijack the original instructions.
• Not measuring: changing a prompt "by eye" without comparing results before and after is like editing code without tests.

Prompt vs prompt engineering

The two terms are often confused. The prompt is the artifact (the text that goes in); prompt engineering is the discipline of designing, testing and optimizing prompts systematically. The difference is the same as between "a query" and "knowing how to design queries".

In short, the prompt is the fundamental interface with generative AI: the point where human intent is translated into an instruction a model can execute. Mastering it, moving from asking for one-off things to designing structured, repeatable and secure instructions, is what turns AI from a curiosity into a business tool.