Prompt Management

Prompt Management lets you store reusable system prompts and apply them to chat requests by sending a prompt ID in a request header. It is designed for teams that want consistent behavior across apps without repeating long system instructions in every request.

One stored prompt becomes your single source of truth-update it once, apply everywhere, and disable it instantly if needed.

Quick Start

Include the X-Prompt-Id header in your chat completion request:

curl https://api.llmapi.ai/v1/chat/completions \
  -H "Authorization: Bearer your_api_key" \
  -H "X-Prompt-Id: your_prompt_id" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-4",
    "messages": [
      {"role": "user", "content": "Hello"}
    ]
  }'

When validated, your stored prompt is injected as the first system message, followed by your original request messages.

How It Works

When you send a request with X-Prompt-Id:

Lookup - LLM API retrieves the prompt by ID from cache or database
Validate - LLM API confirms the prompt exists, belongs to your project, and is enabled
Inject - The stored prompt text is prepended as the first system message
Route - The full message list continues through normal model and provider routing
Respond - The model processes the injected prompt plus your original messages

INFO

Existing client-provided system messages are preserved and appear after the injected prompt.

Errors and Validation

Condition	Status	Code
Prompt ID does not exist	400	`prompt_not_found`
Prompt belongs to another project	403	`prompt_project_mismatch`
Prompt is disabled	400	`prompt_disabled`

WARNING

Prompt data is project-scoped. Cross-project prompt usage is blocked even if someone knows a prompt ID.

Best Practices

One prompt per task family - Create separate prompts for support, coding, extraction, etc. rather than storing multiple unrelated instructions in one prompt.
Use clear, version-like titles - Name prompts descriptively (e.g., "Support Agent v2", "Code Reviewer v1") to make history and intent clear.
Roll out with enable/disable toggles - Disable a prompt instantly if needed instead of deleting it. This lets you revert safely without code changes.
Keep prompt text concise - Shorter prompts are faster to inject and easier to update. Test for regressions after any text changes.

Limitations and Scope

Endpoint support - Prompt injection is currently available for chat completions.
Injection mode - Prompts are always injected as the first system message.
Cache lookup - Lookups use Redis read-through cache with database fallback for high performance.
Cache invalidation - Update or delete operations automatically invalidate the prompt cache.