Prompt Engineering

💬 In Plain Words: The 1-Minute Summary Tool calling in the Model Context Protocol (MCP) is not a one-off button press—it is a complete 6-step conversation loop. Think of an AI model like Claude or a Salesforce Agentforce agent as a smart new hire sitting at a desk. The AI cannot magically reach into your database on its own. Instead, it looks at a menu of approved software tools, picks the right one, fills out a request form, and asks the host application to run it. Once the system hands back the data, the AI reads the result and decides what to do next.

💬 In Plain Words: How AI Reads and Remembers An AI model never reads whole sentences the way a human does. Instead, it chops your text into bite-sized pieces called tokens . Everything the AI can look at during a single conversation—your question, your chat history, any Salesforce records you attach, and the AI's own reply—must fit inside a fixed bucket called the context window (its short-term working memory). The prompt is the complete package you place inside that bucket. When you combine clear instructions with clean business data inside that window, you get accurate, enterprise-grade answers every time. ⚡ Key Points: 1-Minute Summary Large Language Models (LLMs) do not read words directly. They break text down into tokens —small fragments of words, spaces, or punctuation converted into numeric IDs. In standard English, 1 token is roughly 3/4 of a word (so 100 tokens ≈ 75 words, or ~4 characters per token), though code, JSON, and non-Engli...

💬 In plain words: A variable securely stores a specific fact for the rest of an AI conversation. Without variables, you are blindly trusting the AI model to remember things accurately. By using variables, you lock that fact in place so it never drifts, hallucinates, or disappears.

⚡ 1-Minute Summary Agent testing is statistical, not pass/fail. Never assert on exact wording. Assert on which subagent handled the request, which action ran, the decided value, and whether it correctly refused. Measure your noise floor : run the same suite three times unchanged and record the variance.

⚡ 1-Minute Summary Instructions resolve top to bottom. Logic lines run. Prompt lines become English. The model never sees your script. It sees only the finished prompt your logic built. Shorter reasoning instructions produce more accurate agents. Verbosity is not thoroughness. Module map MODULE 2 root: 'What actually happens in one turn?' ├─ 2.1 Anatomy of one turn ├─ 2.2 How the prompt gets built └─ 2.3 Why shorter instructions win Anatomy of One Turn 💬 In plain words: Every message follows the same seven steps: router, before-reasoning, instructions, model, tools, after-reasoning, reply. Learn this sequence and every design decision in Agentforce explains itself. 📌 Example: A Meridian customer says "is my power back yet?". The router picks the outage subagent. Before-reasoning verifies their meter. Instructions run a status action and branch. The model phrases the answ...

💬 In plain words: An instruction tells an AI model what it should do. A filter (or guardrail) controls what an AI model can do. A determined user can always manipulate an AI into ignoring what it "should" do, but nobody can bypass hard-coded system filters that dictate what it "can" do.

💬 In Plain Words: When building enterprise AI agents, you must decide in advance which parts of a conversation the Large Language Model (LLM) is allowed to influence, and which parts are strictly off-limits. Remember this golden rule: An instruction is just a request; a filter is an absolute control. When money, compliance, or eligibility are on the line, you cannot rely on polite instructions in a prompt. You must use hardcoded determinism mechanisms to force certainty.

💬 In plain words: A Prompt Template is simply a saved, version-controlled, reusable AI prompt. Salesforce offers four distinct types. And here is the golden rule: You do not manually type merge field syntax by hand. You write a placeholder word, and then use the built-in Resource Picker to replace it. 📌 Real-Life Example: Field Crew Triage Imagine an energy company using a "Flex" template to help field crews. The template takes a Case ID as input and instantly generates a triage brief. It combines data from the Case object, the attached Asset (Meter), historical faults from a related list, and a Flow that calculates how many times a crew has visited this address before.

💬 In plain words: When you ask an AI a question, it doesn't search a massive database of pre-written answers. Instead, it reads your prompt, breaks it down into mathematical pieces, and predicts the best sequence of words to generate based on patterns it learned during training. It creates the answer from scratch, every single time.

💬 In plain words: When building AI agents in Salesforce Agentforce, before_reasoning and after_reasoning are two optional logic blocks inside a subagent. One runs exactly before the AI thinks, and one runs exactly after. They contain hard logic (code), not prompt suggestions. This is where your absolute, unbreakable rules must live. 📌 Example: A utility company verification agent checks if a caller actually owns the meter in the before_reasoning block. Later, it escalates any refund request above $100 inside the after_reasoning block. Neither of these critical business decisions is left up to the AI language model.

In plain words: There are two action lists inside a subagent in Agentforce, and they serve completely different purposes. One list runs strictly because your logic explicitly called it. The other acts as a menu of tools that the AI model can choose to use if it thinks they are needed. Confusing these two is the most common mistake developers make when writing Agent Script. 📌 Example: Imagine a utility company's agent (like Meridian). It runs an "outage lookup" deterministically every single time by placing it in the actions block. However, it offers a "book a crew visit" tool inside reasoning.actions , so the model only offers to schedule a crew when the context of the conversation actually requires it.

💬 In plain words: When building custom Apex actions for AI (like Agentforce or Einstein Copilot), never just hand the raw record to the Large Language Model (LLM) and ask it to figure out the answer. While the model will usually get it right, its occasional silent failures are impossible to catch. Instead, do the heavy lifting in your Apex code and return a clear, final decision (a verdict) alongside a plain-language explanation to the model. 📌 Example: Instead of passing a raw Service Contract and Case record to the AI to determine a refund, your Apex should process the logic and return specific outputs like: isEligible = true refundAmount = 1240 reason = "Supply was interrupted for 14 hours, which exceeds the 12-hour guarantee." Now, there is absolutely nothing left for the LLM to guess or infer.

💬 In plain words: The Agentforce AI model chooses which action to call by reading its description. It does not look at the action's name, and it does not read the underlying code. Therefore, an action description is not documentation for human developers—it is the strict selection criteria the model uses to choose between a dozen different tools. 📌 Example: Naming an action "Creates an appointment" is incredibly weak. A strong description looks like this: "Books a field crew visit for a specific meter on a chosen date. Returns the booking reference and confirmed date. Use only after checking whether the fault is on the customer's side." 🎬 Real-Life Example: The Action That Never Got Called The Old/Bad Way: A development team built a flawless Apex action to check for system outages. They lazily set the description to: "Outage utility method." Why it failed: The AI model relies purely on description...

💬 In plain words: Every word you add to an AI prompt acts as additional weight for the model to balance. Longer instructions increase the chances that the model gets distracted or ignores core directions. Salesforce explicitly confirms this principle: keeping reasoning instructions concise produces significantly more accurate results. 📌 Example: A Meridian subagent originally used eleven sentences of detailed prompt instructions and consistently misrouted requests. By shortening the instructions to four clear sentences and moving two strict constraints into code logic, the routing accuracy jumped without changing a single downstream action.