Understanding Prompts

Prompts Shape Agent Behavior Prompts are the instructions and context that guide the AI's reasoning. They define what the agent knows, how it should behave, what it should avoid, and how to handle special cases.

What are Prompts?

Prompts in REDA AI are text instructions stored in Custom Prompt records that are linked to Agent Definitions. Unlike the Persona field (which is a single opening system prompt), Custom Prompts provide additional layers of instruction that refine the agent's behavior.

Think of prompts as a set of rules, guidelines, and context that the AI agent consults when processing user requests. Each prompt serves a specific purpose, and multiple prompts can work together to shape the agent's overall behavior.

The Five Prompt Types

Prompt Type	Field Value	Purpose
Persona Override	Persona_Override	Sets or refines the agent's identity, role, tone, and personality. Can override or extend the base Persona field on the Agent Definition.
Custom Instructions	Custom_Instructions	Specific behavioral rules. Examples: "Always use bullet points", "Only discuss real estate topics", "Keep responses under 100 words".
Schema	Schema	Describes Salesforce data structures (field meanings, picklist values, relationships). Reduces the need for tool calls to describe objects.
Action	Action	Instructions for how to perform specific actions. Examples: "When creating a task, always set the priority to High", "Emails should use this signature template".
Guardrail	Guardrail	Restrictions and safety rules. Examples: "Never discuss pricing with free-tier customers", "Escalate salary questions to HR".

Persona Override vs. Persona Field

There's a distinction between:

Persona field on Agent Definition — The base system prompt that defines the agent's core identity and role.
Persona_Override Custom Prompt — Additional prompts that can refine or override aspects of the base persona. Useful for making adjustments without modifying the original Persona field.

In most cases, you'll use the Persona field for the base persona and reserve Persona_Override prompts for refinements or context-specific overrides.

How Prompts are Assembled

When the AI agent processes a user message, REDA AI assembles the system prompt like this:

Base Persona — The Persona field from the Agent Definition comes first.
Custom Prompts, in Sort Order — All applicable prompts follow in a single sequence ordered by Sort Order (lowest first), regardless of prompt type. The type describes what a prompt is for; it does not control its position. To put a Guardrail before an Instruction, give it a lower Sort Order.
User Message — The user's actual message is processed after all prompts.

"Applicable" means a prompt passes every one of these filters at assembly time:

Active — Is Active is checked.
Agent match — the prompt is linked to this agent, or its Agent Definition field is blank, which makes it a universal prompt that loads for every agent (the built-in language directives ship this way).
Channel match — the prompt's Channel selection includes the channel the conversation is running on.
Context match — if the prompt names Context Objects, the current record's object must match.
Not excluded — the prompt is not on this agent's exclusion list (set via Edit Loadout in the Agent Runtime Preview).
Not an exclusive prompt — a prompt with Exclusive Prompt ticked skips this whole list. It is invisible to agent, channel and context matching, and reaches an agent only when an Exclusive Prompt record assigns it.
Fits the token budget — each prompt carries a token estimate (its Complexity Estimate, or 200 when blank). A prompt that would push the running total past the agent's Prompt Complexity Budget (that is its label on the Agent Definition) is skipped for that turn — later, smaller prompts may still load.

Prompt Structure & Content

Field	Type	Description
Prompt Type	Picklist	One of the five types: Persona_Override, Custom_Instructions, Schema, Action, or Guardrail.
Content	Long Text Area	The actual prompt text. This is what gets sent to the AI. Can be very long (thousands of characters).
Is Active	Checkbox	Only active prompts are assembled and sent to the AI. Deactivating a prompt disables it without deleting it.
Sort Order	Number	Controls the position of the prompt in the assembled sequence — across all prompt types. Lower numbers come first.
Agent Definition	Lookup to Agent Definition	Which agent this prompt belongs to. Leave it blank to make the prompt universal — it then loads for every agent in the org.
Channel	Multi-select picklist	Which channels the prompt loads on (Internal, External Widget, Community, API). A prompt selected for Internal and External Widget loads for both.
Context Objects	Text	Optional list of object API names. When set, the prompt loads only when the conversation runs in the context of one of those objects.
Complexity Estimate	Number	The prompt's estimated token size, used to enforce the agent's Prompt Complexity Budget during assembly. Blank counts as 200.

Context Variables in Prompts

Prompts can include context variables that are dynamically replaced with actual values at runtime. Common variables:

{USER_NAME} — The full name of the Salesforce user
{USER_FIRST_NAME} — Just the first name
{USER_EMAIL} — The user's email address
{TODAY_DATE} — Today's date
{AGENT_NAME} — The agent's Display Name
{USER_LANGUAGE} — The user's Salesforce language locale

Example: "Greet the user as {USER_FIRST_NAME}. Today's date is {TODAY_DATE}." becomes "Greet the user as John. Today's date is March 5, 2026."

Common Prompt Examples

Schema Prompt Example

You have access to the following Salesforce objects:

Opportunity:
- StageName (picklist): Prospecting, Qualification, Needs Analysis, Value Proposition, Decision Makers, Perception Analysis, Proposal/Price Quote, Negotiation/Review, Closed Won, Closed Lost
- Amount (currency): Total deal value
- ExpectedRevenue (currency): Forecast revenue
- CloseDate (date): Expected close date

Account:
- Industry (picklist): Technology, Healthcare, Finance, Manufacturing, etc.
- AnnualRevenue (currency): Total annual revenue
- NumberOfEmployees (number): Headcount

Custom Instructions Prompt Example

Follow these behavioral guidelines:
1. Always format lists with bullet points, not numbers
2. Keep responses concise — maximum 3-4 sentences
3. Use friendly, conversational language
4. Never make up information about records; always query to verify
5. Offer to help with the next step in the workflow

Guardrail Prompt Example

DO NOT discuss the following topics:
- Employee compensation or salary details
- Confidential pricing information
- Personal health or medical information
- Credentials or passwords

If asked about these topics, politely decline and offer to escalate to the appropriate department.

Multiple Prompts of the Same Type

You can create multiple prompts of the same type and assign them all to an agent. They are concatenated in Sort_Order sequence. This is useful for:

Modularity — Break complex instructions into multiple focused prompts.
Reusability — Use the same prompt on multiple agents without duplication.
Inheritance Patterns — Create a base prompt that applies to all agents, then agent-specific overrides.
Dynamic Control — Disable/enable prompts without deleting them, or reorder them by changing Sort_Order.

For example, you might have:

Custom_Instructions (Sort_Order 1) — General tone and format rules
Custom_Instructions (Sort_Order 2) — Domain-specific rules for Sales
Custom_Instructions (Sort_Order 3) — Company-specific policies

Deactivating vs. Deleting Prompts

Deactivate (turn off Is Active) — The prompt remains in the database but is not sent to the AI. Use this to temporarily disable a prompt while testing changes.

Delete — The prompt record is permanently removed. Use this only when you're sure the prompt will never be needed again.

Best practice: Deactivate rather than delete, so you can quickly reactivate if needed.

Configure reusable business instructions

Exclusive Prompts let administrators assign reusable instructions to selected agents. Review the Exclusive Prompt flag and Summary on Custom Prompt, then the assignment's Agent Definition, Custom Prompt, Active and Load Mode. Use the intended load mode; On Demand needs the context-prompt tool in the agent's loadout. Record-type brackets use shared scope rules, including in preview and prompt administration. Prompt and tool ordering is deterministic, and runtime preview can use an actual record ID. Updated language instructions follow the message on API/external channels; user placeholders apply only where the authenticated user context is intended. Preserve custom prompts when merging the current defaults.

Frequently Asked Questions

Yes. Leave the prompt's Agent Definition field blank and it becomes a universal prompt that loads for every agent in the org — the built-in language directives ship this way. To share a prompt with only some agents, either duplicate it per agent, or make it universal and exclude it on the agents that shouldn't use it (Edit Loadout in the Agent Runtime Preview).

The AI will attempt to follow both instructions. However, contradictions can lead to unpredictable behavior. Keep prompts consistent and test carefully. If you find contradictions, use Sort_Order to prioritize one over the other, or consolidate them into a single prompt.

Prompts are plain text, not HTML. You can use markdown-style formatting (*, -, # for lists and headers) if the AI interprets it, but the safe approach is plain text. The AI will handle formatting in its response independently.

The prompt Content field is a Long Text Area holding up to 20,000 characters. In practice, keep prompts under 10,000 characters per prompt for clarity and to avoid excessive token consumption. If you need extensive guidance, split it across multiple prompts.

Next Steps

Learn how to create and edit prompts for your agents, or explore prompt best practices to optimize your agent's responses.

Next: Creating & Editing Prompts →