Custom Prompts Per Agent
Agent-Specific Prompt Architecture
In REDA AI, Custom Prompts are stored as records with a lookup relationship to Agent Definition records. This design ensures that:
- Each agent has its own isolated set of prompts (plus any universal prompts — prompts with no agent set — which load for every agent)
- Prompts for one agent do not affect other agents
- You can create completely different prompt strategies for different agents
- Agents can coexist in the same org without interfering with each other
This is a fundamental design principle of REDA AI — agents are independent units with their own configuration, including prompts.
Multiple Agents with Different Prompts
A common scenario is having multiple agents for different departments, each with tailored prompts:
| Agent | Department | Key Prompts |
|---|---|---|
| Sales Assistant | Sales | Persona: Enterprise SaaS specialist; Schema: Opportunity, Account fields; Guardrail: No pricing discussions; Action: How to create tasks |
| Support Bot | Customer Support | Persona: Empathetic support specialist; Schema: Case, Contact fields; Custom Instructions: Offer solutions before escalating; Guardrail: Protect customer data |
| HR Assistant | Human Resources | Persona: Professional HR advisor; Schema: Employee fields (limited); Guardrail: Confidentiality rules; Action: Direct to HR portal |
| Finance Bot | Finance | Persona: Analytical finance expert; Schema: Expense, Invoice fields; Guardrail: Compliance rules; Custom Instructions: Cite policy numbers |
Each agent has a completely different tone, instruction set, and scope based on its department's needs.
Sharing Common Instructions Across Agents
When the same instruction should apply to more than one agent, you have two mechanisms — and the first one means you maintain the text exactly once:
- Make it a universal prompt — Create the Custom Prompt and leave its Agent Definition field blank. A universal prompt loads for every agent in the org (the built-in language directives ship this way). Channel and Context Objects scoping still apply, so a universal prompt can target only certain channels or record contexts.
- Opt out the exceptions — If one or two agents shouldn't use a universal prompt, exclude it on just those agents with Edit Loadout in the Agent Runtime Preview — the prompt stays universal for everyone else.
- Assign it exclusively when sharing is partial — If only a few specific agents need the instruction and most should not have it, do not clone the prompt. Tick Exclusive Prompt on it and assign it to those agents with Exclusive Prompt records, described below. One record, several agents, no copies to keep in sync.
Prefer the universal prompt whenever the instruction is genuinely org-wide — one record to edit, no copies to drift.
Exclusive Prompts: one prompt, several agents, on purpose
A prompt normally finds its agents by itself — through the Agent Definition field, the Channel and the Context Objects. That is right for most prompts and wrong for one case: a procedure that several particular agents need and no others should pick up.
Tick Exclusive Prompt on the Custom Prompt and the usual rules stop applying to it entirely. It is then invisible to every agent until an Exclusive Prompt record assigns it to one — and each assignment decides for itself how the prompt loads. Leave the checkbox unticked and the prompt behaves exactly as prompts always have.
| Load Mode | When the prompt is included |
|---|---|
| Always (default) | Every turn — and it is the answer for turns that have no record context at all. |
| On Context | Only when the turn is on a matching object, and matching record type where one is set. |
| On Demand | Not loaded at all. The prompt is listed in the agent’s fetchable catalogue so the agent can request it by name when the situation calls for it — the cheapest option, since the text costs nothing until it is actually needed. |
On Context needs somewhere to match. Saving an assignment as On Context when the prompt has no Context Objects is blocked: “On Context needs the prompt to say which objects it applies to. Set Context Objects on the prompt (for example Account, or Account[Customer]), or change Load Mode to Always.”
Two more things worth knowing about the assignment record. Active lets you suspend one assignment without deleting it — an inactive assignment is ignored entirely. And the assignment is master-detail to the agent but only a lookup to the prompt, with delete restricted: removing an agent cleans up its own assignments, while a prompt that other agents still depend on cannot be deleted out from under them.
Organization Pattern: Base Agent + Specialized Clones
A best practice is to create a "base" agent with a well-configured set of prompts, then clone it for specialized use cases:
-
Create Base Sales Agent
Create an agent with Display Name "Sales Agent - Base Template" with standard prompts for sales tasks.
-
Clone for Each Region
Clone the agent to create "Sales Agent - North America", "Sales Agent - EMEA", "Sales Agent - APAC".
-
Customize Each Clone
For each regional clone, customize the Persona and prompts with region-specific language, rules, or focus.
-
Deploy Each Independently
Each regional agent is now independent and can be deployed to different Salesforce environments or pages.
This approach scales well — you maintain a consistent base architecture while allowing customization.
Org-Wide Prompts: Leave the Agent Blank
REDA AI does support organization-level prompts: a Custom Prompt whose Agent Definition field is blank is universal and loads for every agent automatically — the built-in language directives are shipped exactly this way. Agent-linked prompts stay isolated to their agent; universal prompts are the deliberate escape hatch for org-wide rules.
Two controls keep universal prompts safe: they respect Channel and Context Objects scoping like any other prompt, and any individual agent can opt out of one via Edit Loadout in the Agent Runtime Preview.
Different Tones for Different Departments
One common use case is having different tones for agents serving different departments:
"You are an enthusiastic Sales Account Executive assistant. Your tone is upbeat, encouraging, and ambitious. You help sales reps close deals, move opportunities forward, and build customer relationships. You celebrate wins and help reps stay focused on their pipeline."
"You are an empathetic Customer Support specialist. Your tone is professional, patient, and solution-focused. You help support reps resolve customer issues quickly, reduce frustration, and delight customers. You prioritize understanding customer needs before recommending solutions."
"You are a professional Human Resources advisor. Your tone is respectful, confidential, and policy-aware. You help HR teams manage employee inquiries, ensure compliance with company policies, and maintain confidentiality. You always escalate sensitive matters to the HR director."
Each agent's tone and approach is distinctly different, shaped by their prompts.
Domain-Specific Context and Schema Prompts
Different agents work with different Salesforce objects. Customize Schema prompts for each agent's domain:
"Primary objects: Opportunity (with StageName, Amount, ExpectedRevenue, CloseDate), Account (Industry, AnnualRevenue, NumberOfEmployees), Contact (Title, Department, Email, Phone). Relationships: Accounts have multiple Contacts and Opportunities; Contacts belong to Accounts."
"Primary objects: Case (with Status, Priority, Origin, Category), Contact (Email, Phone, Account, SupportTier), Account (CurrentServiceLevel). Relationships: Cases belong to Contacts and Accounts; Support cases flow through resolution workflows."
Each agent's Schema prompts describe only the objects and fields relevant to that agent's work.
Guardrails: Department-Specific Restrictions
Guardrail prompts often differ by department:
"Do not make commitments on delivery timelines without verifying with the product team. Do not discuss competitor features in detail. Always escalate pricing objections to management."
"Keep all employee personal information confidential. Never discuss salary or compensation publicly. Escalate disciplinary matters to the HR director. Protect applicant privacy per legal requirements."
Each department has different compliance and policy requirements reflected in guardrails.
Testing Agent-Specific Prompts
When testing prompts, ensure you're testing the right agent:
- Deploy the specific agent whose prompts you're testing
- Don't assume that a prompt on Agent A will work the same on Agent B
- Test department-specific behavior with realistic scenarios for that department
- Verify that guardrails specific to each department are effective
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 universal, loading for every agent in the org automatically. To share with only some agents, clone the record per agent (changing the Agent lookup), or make it universal and exclude it on the agents that shouldn't use it via Edit Loadout.
You can't delete the agent while its prompts exist — Salesforce blocks the deletion until the related Custom Prompts (and Context Fields and Suggestion Chips) are removed first. Deactivating the agent instead is the best practice: everything stays intact for reactivation later.
While REDA AI doesn't have shared personas at the database level, you can create nearly identical Persona prompts on different agents. Use consistent language and tone to create a unified experience across related agents.
Use naming conventions in the Prompt Name field to indicate which agent or department the prompt is for. For example, "Sales - Schema: Opportunity Fields" or "Support - Guardrail: Customer Privacy". This makes it easy to find and manage prompts.
Next Steps
Learn about context variables to personalize prompts across different agents, or explore prompt best practices for optimizing each agent's performance.