Prompt Template Gallery
What the Gallery Is
The gallery is a small collection of hand-written starter prompts maintained inside the platform. Each template is a complete, working prompt body — not a fill-in-the-blank skeleton. You pick the template that matches the kind of prompt you want, edit anything that doesn’t fit your agent, and save the result as a Custom Prompt record.
Templates are not custom-metadata records you can edit; they’re part of the package. Once dropped into the Prompt Builder, the resulting Custom Prompt is yours to change however you like.
Template Categories
Templates are grouped by Prompt Type. Each category exists for a different shape of work:
| Type | What it does | Templates available |
|---|---|---|
| Custom Instructions | The agent’s general response rules — tone, length, format, guardrails, what to do when tools fail. | Three: a compact set, a detailed set, and a customer-facing set. |
| Action | How the agent should propose writes — tasks, events, emails, record updates — and the action-confirmation contract. | Three, covering common action surfaces. |
| Schema | How to describe a Salesforce object’s fields and relationships to the agent so it picks the right ones. | One generic-object starter that you adapt to your custom objects. |
| Persona Override | A targeted persona shape for an agent — tone, voice, scope, what to refuse. | One general-purpose persona starter. |
See Custom Prompts per Agent for the full list of prompt types and how each contributes to the assembled system prompt.
Custom Instructions Templates
Three templates cover the common positions on the rule-density spectrum:
Core Rules — Compact
The leanest set of rules. About 150 tokens. Use it when token budget is tight and you trust the model to do the right thing on most details. Covers: tools-first behaviour, output length cap, formatting markers (bold labels, tables for 3+ records), the link and alert syntaxes the chat UI expects, suggestion chips at the end, and the {TODAY_DATE} placeholder.
Core Rules — Detailed
The same baseline plus explicit formatting and guardrail sections. About 300 tokens. Reach for this when you want the agent to apply consistent currency formatting, date formatting, and explicit guardrails (no fabrication, no record modifications without confirmation, field-level security (FLS) and sharing respected). Best for production agents serving demanding internal users.
Customer-Facing Rules
Polite, external-safe rules. About 200 tokens. Tone is friendly and professional; the prompt explicitly forbids exposing internal Ids, object names or technical details. Use it as the starting point for any external-widget or community-portal agent.
Action Templates
Three templates describe how the agent should structure write proposals so the platform can render confirmation cards and execute the actions cleanly. The shape is consistent across templates — a structured proposal line the agent emits, which the platform parses and shows to the user. Differences between the three templates are in scope: a comprehensive set covering tasks/events/emails/record updates; a leaner set for read-and-update agents; and a notification-focused set for nudge-style workflows.
Action templates pair naturally with the built-in Action tools listed in the Built-in Tools Reference. The template doesn’t replace the tool definitions — it tells the agent how to propose using them.
Schema Templates
The Schema starter is intentionally generic. The platform ships one template that walks the agent through how to describe an object’s fields, relationships and picklists; you adapt it to your specific custom objects.
Why isn’t there a per-object template library? Because the same package installs into orgs with completely different data models. A canned schema prompt for a domain object you don’t have would be noise. The starter teaches the shape; you fill in the substance.
Schema prompts shine when the same object is queried often. A well-written schema prompt can cut several tool calls (describe_object, get_picklist_values, get_record_types) per conversation by giving the agent the relevant facts upfront. See Configuring Tools for the trade-off.
Persona Override Template
A general-purpose persona shape covering tone, scope, escalation rules and refusals. Use it as a checklist when you write a custom persona; it ensures you don’t forget to specify the things that matter most:
- Who the agent is and who it’s talking to
- What it can and can’t do
- What tone fits the channel
- What to refuse, and how to refuse politely
- Where to escalate when it can’t help
Working With a Template
-
Open Prompt Builder
From the REDA AI Admin tab or directly from a Custom Prompt record.
-
Pick the prompt type
The template gallery filters to that type so you only see relevant starters.
-
Insert the template
The template content drops into the editor. The estimated token count appears next to it.
-
Edit for your agent
Tighten the rules, change tone, swap placeholders for your real values, remove anything that doesn’t fit.
-
Save
The save creates a regular Custom Prompt record. The Quality Analyzer runs automatically — you’ll see a quality score on the record.
-
Optimise if needed
If the analyser flags issues, run the Prompt Optimizer to clean them up while keeping the meaning.
Why These Templates Score Well
The shipped templates are built against the same heuristics the Quality Analyzer applies:
- Headers (
## RULES,## FORMATTING,## GUARDRAILS) keep them from triggering the UNSTRUCTURED and NO_SECTIONS checks. - Bullet lists make rules easy to extend without restructuring.
- No verbose filler — the templates use short imperatives, not “please make sure to…”.
- No redundant assertions — nothing the model already does by default is restated.
- Sized to fit the agent’s typical token budget without dominating it.
The templates are deliberately short. Most agents don’t need more.
What the Templates Do Not Cover
- There is no Guardrail template. Guardrail is one of the five prompt types, and the gallery ships nothing for it — there is not even a Guardrail tab. That is deliberate: a guardrail is a statement of what your organisation will not allow, so a generic one would be worse than none. The best-practices guide shows how to write one, and the Custom Instructions templates include guardrail-shaped clauses you can lift.
- They’re not industry-specific. No real-estate, sales, support or finance vocabulary — the platform is horizontal. Bring your own domain language.
- They’re not localised. All template content is in English. The Multi-Language mechanism handles translation at runtime; the templates themselves don’t need to be translated.
- They don’t cover every prompt type in depth. One template per niche type (Schema, Persona Override) is the gallery’s scope. For deeper guidance, write the prompt and run it through the analyser and optimiser.
- They don’t reference your custom objects or fields. Generic on purpose — the prompts you save are where the org-specific detail goes.
Best Practices
Start with a template, then trim. Edit toward your specific agent rather than building from scratch. Most of the rule-density work is already done.
Keep one Custom Instructions prompt per agent. Two competing instruction prompts in the same agent is a common conflict source. Pick the right starter, customise once, leave the others off the agent.
Customise the customer-facing template before going live. The shipped customer-facing rules are safe defaults but you almost certainly want to add a sentence or two about your brand voice.
Treat the schema starter as a teaching example. Read it once, then write your own schema prompts using the same shape against your real custom objects. The starter is not meant to be saved verbatim.
Run the analyser after editing. Edits can introduce verbose filler or reorganise sections. The score shifts; check it before you call the prompt done.
Changes included in 1.16000
When copying an older tool example, use the current signature: send_email uses to, and update_record uses field_updates. Task and event tools support field_values for additional permitted fields. Verify the active tool’s current input schema before using an older saved template.
Frequently Asked Questions
The shipped templates are part of the package. You can’t add to or edit the gallery itself. What you can do is build a library of saved Custom Prompt records in your org and reuse them across agents — which gives you a customer-specific gallery.
No. The templates are starting points. You can write any Custom Prompt from scratch and skip the gallery entirely.
Yes. They’re shipped with the package and are identical for every install. Differences between orgs come from the Custom Prompts you save after editing.
It includes explicit formatting rules and guardrails. Where the compact template trusts the model to format and guardrail by default, the detailed template spells out currency formatting, date formatting, the no-fabrication rule and field-level security (FLS) respect explicitly. Use the detailed version when you’ve seen the compact version produce inconsistent output.
The starter teaches the shape that works well: heading, the key fields the agent should know about, picklist values it should use, relationships to follow. Once you see the pattern on a generic example, applying it to your custom objects is mechanical.
The Optimizer runs against saved Custom Prompts, so you save first, then optimise. The shipped templates already score well, but if you’ve made significant edits, the Optimizer can tighten the result.
Next Steps
Pick a template and head to the Creating & Editing Prompts guide, or run the Quality Analyzer on prompts you’ve already written to see how they compare.