Skip to main content

​
What are prompts?

In Basalt, a prompt is a reusable, versioned template that describes how your AI should behave in a given context. Instead of hard‑coding strings like “You are a support bot…” all over your codebase, you define prompts once in Basalt, give them a stable slug, and reference them from your applications via the SDK or API. A prompt typically combines:
  • System behavior – role, tone, and constraints (e.g. “You are a senior support agent…”)
  • Task instructions – what the model should do (e.g. “Summarize this ticket in 3 bullet points…”)
  • Examples – optional few-shot examples to steer behavior
  • Variables – placeholders like {{customer_message}} filled at runtime
  • Model configuration – provider, model name, and parameters (temperature, max tokens, etc.)
Prompts become first-class, shareable building blocks you can reuse across:
  • User-facing features (assistants, copilots, search, recommendations)
  • Internal tools (support, operations, data quality checks)
  • Background workflows (cron jobs, pipelines, agents)

​
Prompts Management

Basalt’s Prompts API lets you manage these templates like code:
  • Prompts are versioned so you always know what is running.
  • You deploy versions to environments using tags (for example production, staging, latest).
  • You can evolve prompts safely over time while keeping older versions available for rollback or comparison.
With the Prompts API you can:
  • Create and update prompts as your product or use cases evolve
  • Track versions to understand which behavior is live in each environment
  • Deploy using tags instead of hard-coding version numbers in your application
  • Roll out changes gradually by pointing different services or environments at different tags
  • Attach model settings directly to prompts so configuration travels with the template

​
Overview

At a high level, the Prompts API enables you to:
  • Store prompts centrally with full version history
  • Deploy prompts with tags (e.g. production, staging, latest)
  • Substitute variables dynamically using Jinja2 templates
  • Cache prompts for performance with graceful fallback
  • Manage model configurations alongside prompt content
In the rest of this section you’ll see:
  • How to define prompts and variables in Basalt
  • How to fetch prompts from the SDK by slug, tag, or version
  • How to safely update and roll back prompts with versioning
  • How to test and iterate on prompt changes without risking production