Prompt Builder and Agent Builder in practice

Prompt Builder and Agent Builder are configuration surfaces, so they deserve the same review discipline as a validation rule.

1. VideoGetting an org with generative AI enabledPreview6 min
2. VideoTemplate types, and choosing by where the output lands6 min
3. VideoPrompt Builder against Agent Builder: two tools, one model6 min
4. VideoReading Salesforce’s own prompt gallery for the trick, not the copy6 min
5. ReadingThe org you need, and the trial path if you don’t have one5 min
6. LabLab: build your first template and run it once14 min

The fields you send are the fields the model can leak. Choose them the way you would choose what goes in a report.

1. VideoGrounding: picking the fields, and the ones you must not send6 min
2. VideoMerge fields, related lists, and flow-based grounding6 min
3. VideoThe field you added because it was there6 min
4. VideoPII minimisation inside a grounded template6 min
5. ReadingA grounding checklist before you save5 min
6. LabLab: build a grounded summary template14 min

Instructions written for the happy path fail the first time a user does not follow it. Write for the second user, not the first.

1. VideoWriting instructions that hold under an unusual record6 min
2. VideoThe user who goes off the happy path6 min
3. VideoNegative instructions: what the model must never do6 min
4. VideoReal-data-only: no placeholder names, no round invented figures6 min
5. ReadingDate grounding, so “this year” means this year5 min
6. LabLab: break your own instructions, then fix them14 min

A topic is a boundary and an action is a capability. Getting the boundary wrong is how an agent answers a question it should have escalated.

1. VideoTopics, classification, and the boundary between them6 min
2. VideoStandard actions, Flow actions and Apex actions6 min
3. VideoWhat each action is allowed to write, decided before it is built6 min
4. VideoInput and output types that do not drift6 min
5. ReadingNaming topics and actions so the model picks correctly5 min
6. LabLab: define two topics and the actions under each14 min

Actions that write need more discipline than actions that read. This module wires both, and adds the guardrail that keeps a raised exception from killing the whole run.

1. VideoWiring a Flow action with input and output variables6 min
2. VideoWiring an Apex-invocable action the same way6 min
3. VideoA safe-tool wrapper: never let a raised exception kill the run6 min
4. VideoEvery write wrapped in a required confirmationThe user gets an approve button. Nothing writes without one, no exceptions for convenience.6 min
5. ReadingAuth errors marked non-retryable, with a hint to stop and ask5 min
6. LabLab: build an agent with two topics and one write action14 min

Every one of these recurs across four production agents. That repetition is the evidence they are not incidental.

1. VideoQuery coaching inside the error hint6 min
2. VideoThe real-data-only instruction, and what happens without it6 min
3. VideoDate grounding on every run6 min
4. VideoBulk edits as one approval, not N6 min
5. VideoNever claim success from memory: re-query by id after a write6 min
6. ReadingFour agents, one guardrail list: what they all share5 min
7. LabLab: add a safe-tool wrapper to a failing action14 min

What the agent must not answer is as much a design decision as what it may. Then learn to read the trace that shows why it chose what it chose.

1. VideoEscalation: what the agent must refuse to answer6 min
2. VideoReading a trace of an agent turn6 min
3. VideoWhy the agent chose that action, and why it chose the wrong one6 min
4. VideoAuth errors reaching the user, not swallowed by a retry loop6 min
5. ReadingDrafting the refusal message so it does not read as a bug5 min
6. LabLab: read three traces and explain each decision14 min

The gap between a demo and production is a test set. This module builds one, runs it, and grades it without grading the wording.

1. VideoBuilding a test set before you shipTwenty cases with expected outcomes, including five that must be refused. Run it on every change.6 min
2. VideoTwenty cases, five of which must be refused6 min
3. VideoThe accuracy drop from test to production, and where it comes from6 min
4. VideoRunning the case set on every change to the instructions6 min
5. ReadingGrading a transcript without grading the wording5 min
6. LabLab: build and run a twenty-case test set15 min
7. QuizCheck: Prompt Builder and Agent Builder9 min