Skip to main content

Prompt store

Updated on 3 November 2026.

Store reusable system prompts centrally, then reference one or more of them and add inline instructions at request time. You can also give an app one universal system prompt that the gateway adds to every request.

Configure this control​

App prompt definitions stay editable in both modes. Choose Classic settings for app prompt setup, or Policy Engine for organization-wide prompts and scoped enforcement. Only the governed enforcement and universal-prompt settings follow the engine switch. See What happens to classic settings.

Going further with the Policy Engine​

The Prompt Store & Enforcement card in Policy Engine > LLM Gateway holds three things: the organization-wide Global Prompt Store, the Require store prompt policy and the Universal system prompt policy. When the engine is on for the LLM Gateway, the app's store-prompt enforcement and universal system prompt freeze, and the card's policies decide them. App prompts themselves stay editable. See What happens to classic settings.

Global Prompt Store​

The Global Prompt Store is one prompt library for your whole organization, reusable by every LLM Gateway app. Open Govern > Policy Engine > LLM Gateway > Controls > Prompt Store & Enforcement, then choose its Prompt Store tab.

The Prompt Store tab lists every prompt with its ID, variables and content. Search by ID or content, or use Add prompt, Edit and Delete. IDs and {{variable}} rules are the same as for app prompts. Changes apply immediately across the organization: they are not part of the policy draft and do not require publishing.

App Prompt StoreGlobal Prompt Store
WhereApp Configure > PromptsPolicy Engine > LLM Gateway > Prompt Store & Enforcement > Prompt Store
ScopeOne appEvery app in the organization
Reference in requestsquilrai-prompt-store-<id>quilrai-prompt-store-<id>
VariablesX-Prompt-Variables headerX-Prompt-Variables header

How the two stores relate:

  • The global list also shows the prompts of your active apps. When apps share an ID, the newest app version is the default and conflicting versions get alias IDs.
  • Saving an app-derived ID in the global store creates an organization-level version that takes priority over the app copies.
  • Deleting an app-derived ID in the global store hides it from the global list.

Require store prompt​

Use a recipe or the add menu in Require store prompt to add a configuration with Require store prompt set to Required, Not required (an exception that must outrank the matching requirement) or Leave as is. It applies on chat, responses and vertex. The highest-priority matching configuration wins; Required wins ties. Edits join the shared draft and apply once you publish a revision.

Scenarios:

  • Production only. Require a store prompt when request metadata marks the environment as production, and leave development free to experiment.
  • Tenant-wide with exceptions. Require it for Everyone, then set Not required for one Application, App tag or Smart group.
  • Per model or provider. Require it only for a Requested model or Provider.
require_approved_system_promptsrequest

runs on requestpriority 600

WhenRequest metadata . environmentisproduction
Then
Require Prompt Store system promptrequired

Universal system prompt policies​

With the engine on, the universal system prompt comes from policies instead of each app's setting. Open Universal system prompt in the Prompt Store & Enforcement panel and start from a recipe or its add menu. Choose who it applies to, then set Universal system prompt to:

  • On, with the prompt from either:
    • Write the prompt: type it in Prompt text. The counter shows its length against the 8,192-character limit.
    • Use a Global Prompt Store prompt: pick it in Global Prompt Store prompt, for longer prompts or one prompt shared by several configurations. Manage prompts in the Prompt Store opens the Global Prompt Store.
  • Off: an explicit exception at a higher priority than the matching On configuration.

The card shows each configuration as On · 1,234 characters, On · Global prompt support-persona or Off. Editing one reopens the same form, or the full editor for a configuration the form cannot show.

The gateway adds the winning prompt exactly as described in What the model receives, on chat (including Anthropic Messages), responses, vertex and bedrock requests.

How it resolves:

  • One prompt per request. The highest-priority matching configuration wins. The form sets a priority from the scope, and the form then moves to the next priority no other universal system prompt holds. Two configurations at the same priority with different prompts cannot be published unless they apply to different applications.
  • No matching configuration, no prompt. The app's own universal system prompt does not apply while the engine is on.
  • Global Prompt Store prompts are live. The gateway reads the prompt on each request, so editing it in the Global Prompt Store applies immediately, without publishing. It is sent as written: {{variables}} are not resolved.
  • A deleted prompt is skipped. Requests go to the model without a universal system prompt, and the request log records the reason prompt_not_found. The form flags a prompt the Global Prompt Store no longer holds, and validation warns about it.
  • Request logs name the source. The log records that the prompt came from a policy and, for a Global Prompt Store prompt, its ID. The prompt text is never logged.

When you turn the engine on, each app's universal system prompt becomes a configuration scoped to that app. A prompt longer than 8,192 characters cannot move automatically: the conversion review flags it, and you add it to the Global Prompt Store and reference it from a configuration.

Scenarios:

  • One persona for everyone. Turn it on for Everyone with a Global Prompt Store prompt, and set Off for the one Application that brings its own.
  • Per team. Give a Smart group its own prompt at a higher priority than the Everyone configuration.
  • Per model. Add model-specific instructions for one Requested model.
support_team_personarequest

runs on requestpriority 600

WhenSmart groupsincludes (ignoring case)Support
Then
Universal system promptOn · Global prompt support-persona

How it works​

Prompt Stored
ID: code-reviewer
"You are a {{tone}} reviewer"
API References It
system: quilrai-prompt-store-code-reviewer
vars: {tone: "formal"}
QuilrAI Resolves
"You are a formal reviewer"
Sent to LLM ✓
QuilrAI
  1. Create - Store a prompt with a unique ID (e.g., code-reviewer)
  2. Reference - Use it as the system message content: quilrai-prompt-store-code-reviewer
  3. Gateway Resolves - The gateway resolves the prompt and sends the full text to the LLM

Combining prompts and instructions​

A system message is not limited to a single reference. The gateway scans it for quilrai-prompt-store-<id> reference tokens and replaces each one in place with that prompt's resolved content, leaving any other text exactly where you wrote it. So one system message can:

  • Reference several prompts - list multiple quilrai-prompt-store-<id> tokens and the gateway combines their resolved content in the order written.
  • Mix in your own instructions - add freeform text around the references to extend the stored prompt for a single request, without editing the stored prompt itself.

System message your app sends:

quilrai-prompt-store-startup-advisor

Be concise, practical, and honest. Always answer in English.

Stored prompt startup-advisor:

You are an experienced startup advisor, venture capitalist, product strategist, and entrepreneur.

Your role is to challenge assumptions, identify risks, and provide actionable recommendations.

Resolved prompt sent to the LLM:

You are an experienced startup advisor, venture capitalist, product strategist, and entrepreneur.

Your role is to challenge assumptions, identify risks, and provide actionable recommendations.

Be concise, practical, and honest. Always answer in English.

Referencing multiple prompts works the same way - each token resolves independently and the surrounding layout is preserved:

quilrai-prompt-store-base-policy
quilrai-prompt-store-code-reviewer

Only review the security-sensitive files in this diff.
tip

References and inline text are combined from top to bottom in the order they appear. Put foundational prompts first and request-specific instructions last so the model reads them in a natural order.

Universal system prompt​

A universal system prompt is one prompt per app, with no Prompt ID, that the gateway adds to every request. Callers do not reference it, so it reaches the model even when an app sends its own system prompt or none at all.

Open Govern > Policy Engine > LLM Gateway > Controls > Prompt Store & Enforcement and add a rule under Universal system prompt. Choose who it applies to and whether to write prompt text or reference a Global Prompt Store prompt, then Save to draft and Review & publish. See Universal system prompt policies for precedence and prompt sources.

What the model receives​

The gateway adds the universal system prompt before the system prompt your app sends, separated by a blank line. A request with no system prompt gets the universal system prompt as its system prompt. Prompt Store references in your app's system message are resolved first, so the universal system prompt sits in front of the resolved text.

Universal system prompt:

You are Acme's support assistant. Never share internal URLs.

System message your app sends:

Answer in Spanish.

Sent to the LLM:

You are Acme's support assistant. Never share internal URLs.

Answer in Spanish.

When the request has no system message, the LLM receives only You are Acme's support assistant. Never share internal URLs. as its system prompt.

Where it goes in each format​

Request formatWhere the universal system prompt is added
Chat CompletionsMerged into the first message when it is a system or developer message, otherwise added as a new first system message
ResponsesPrepended to instructions, which is set when absent
Anthropic MessagesPrepended to system: in front of a string, or as a new first text block when system is a list of blocks
Gemini / Vertex generateContentThe first part of systemInstruction, which is created when absent. Not applied when the request uses cachedContent, because Gemini does not accept a system instruction with cached content
Bedrock Converse and ConverseStreamThe first system block
Bedrock InvokeModelClaude models (system), Nova models (system blocks) and OpenAI models (messages). Other model families are sent unchanged

It is not applied to embeddings, legacy completions, images, moderations, Realtime, Assistants or Anthropic token counting.

How it behaves​

  • Sent as written. Prompt Store references and {{variables}} inside the universal system prompt are not resolved.
  • Applied after guardrails. The gateway adds it after guardrails run and Prompt Store references are resolved, right before the request goes to the provider, so guardrails never scan or redact it. It is added to the request in the format your client sent, so it also works with any-to-any formats.
  • Not in request logs. Logs keep the request exactly as your app sent it.
  • Counts toward token limits. Its length is part of the token estimate for rate and token limits.
  • Fresh cache entries. Changing it starts fresh entries in the app's response cache.
  • Covered by leakage checks. The System Prompt, Tool & Model Leakage category checks responses against it, along with the rest of the system prompt.
  • Enforcement is unchanged. Enforce system prompts from store checks only the system prompt your app sends.
  • Governed by the Policy Engine. When the engine is on for the LLM Gateway, policies decide the universal system prompt and this setting freezes. See Universal system prompt policies.

To set it from automation, use the Management API.

Template variables​

Prompts support {{variable}} placeholders. Pass values via the X-Prompt-Variables header, keyed by the reference each set of values belongs to.

Prompt template (code-reviewer):

You are a {{tone}} code reviewer for {{language}}.

X-Prompt-Variables header:

{"quilrai-prompt-store-code-reviewer": {"tone": "formal", "language": "Python"}}

Resolved prompt sent to LLM:

You are a formal code reviewer for Python.

When a system message references several prompts, give each one its own entry - the gateway applies each variable set only to its matching reference:

{
"quilrai-prompt-store-code-reviewer": {"tone": "formal", "language": "Python"},
"quilrai-prompt-store-base-policy": {"region": "EU"}
}

Enforce system prompts​

Enforce system prompts from store, in the app's prompt settings, ensures every request's system message includes at least one managed Prompt Store reference, so no request runs without a reviewed base prompt.

ModeBehavior
EnabledEvery system message must contain at least one valid Prompt Store reference (quilrai-prompt-store-<id>). You can list multiple references and add your own inline instructions around them - the request is accepted as long as a valid reference is present. A system message with freeform text but no valid reference, or no system message at all, is rejected with a 400 (system_prompt_not_from_store).
Disabled (default)Both stored references and fully freeform system prompts are accepted.

This applies uniformly across Chat Completions, Anthropic Messages (both the top-level system field and any system-role messages), Vertex/Gemini, and the OpenAI Responses API. Useful when every system prompt should build on a reviewed base from the Prompt Store while still allowing per-request instructions.

The universal system prompt does not count as a reference. Enforcement checks only the system prompt your app sends, so with enforcement on, a request without a valid reference is still rejected even when the universal system prompt is on.

Code examples​

OpenAI​

from openai import OpenAI

client = OpenAI(
base_url='https://guardrails-usa-2.quilr.ai/openai_compatible/',
api_key='sk-quilr-xxx'
)

response = client.chat.completions.create(
model='gpt-4o-mini',
messages=[
{'role': 'system', 'content': 'quilrai-prompt-store-code-reviewer'},
{'role': 'user', 'content': 'Review this code'}
],
extra_headers={
'X-Prompt-Variables': '{"quilrai-prompt-store-code-reviewer": {"tone": "formal", "language": "Python"}}'
}
)

Anthropic​

import anthropic

client = anthropic.Anthropic(
base_url='https://guardrails-usa-2.quilr.ai/anthropic_messages/',
api_key='sk-quilr-xxx'
)

message = client.messages.create(
model='claude-sonnet-4-5',
max_tokens=1024,
system='quilrai-prompt-store-code-reviewer',
messages=[
{'role': 'user', 'content': 'Review this code'}
],
extra_headers={
'X-Prompt-Variables': '{"quilrai-prompt-store-code-reviewer": {"tone": "formal", "language": "Python"}}'
}
)