TechPrufer
Get started
Spotify
Spotify logoSpotify · Feature trail

AI-Assisted API Interaction

1 post · Apr 2026 – Jun 2026

What this trail covers

Spotify is enhancing its AI-Assisted Developer Productivity Platform by integrating LLM evals with online experiments. This creates a feedback loop where LLM evals, acting as automated judges for quality and relevance, precede experiments. The system calibrates LLM judges against real user outcomes from A/B tests, improving the accuracy of evals over time and raising the hit rate of experiments by filtering out less promising candidates. This approach aims to reduce the cost of experimentation by enabling natural language interfaces for complex APIs, such as the Spotify Ads API, through Claude Code plugins. These plugins translate natural language requests into sequences of API calls, leveraging OpenAPI specifications and OpenAPI Links to manage entity hierarchies and workflows. The system also includes pre-flight validation and transparent curl command generation for user auditability and control.

All Spotify trails
  1. Encoding Your Domain Expert: The Context Layer Behind Spotify's Data Assistant | Spotify EngineeringThis post introduces the 'context layer' for Spotify's AI data assistant, which goes beyond simply providing schemas to LLMs. It details the 'cluster model' comprising datasets, vetted question-SQL pairs ('Pairs'), and additional business context ('Docs'), all curated by domain experts. It highlights the importance of human judgment in selecting high-quality training examples, showing that only 12.5% of automatically generated pairs were accepted by curators. The post also describes a 'cluster health score' mechanism to monitor the freshness and validity of the context, and how user conversations and feedback close the loop for continuous improvement.
  2. Building a Natural Language Interface to the Spotify Ads API with Claude Code Plugins | Spotify EngineeringThis post details the development of a Claude Code plugin for the Spotify Ads API, enabling natural language interaction for campaign management. It introduces a Markdown-based plugin architecture with skills, agents, and hooks, and explains the choice of CLI + OpenAPI Spec over MCP due to API complexity and the need for transparency. The post highlights the use of OpenAPI Links to encode entity hierarchies and workflows, allowing the agent to decompose complex requests into sequential API calls. It also discusses the role of the OpenAPI spec as both documentation and validation source, and the use of curated documentation alongside the raw spec for agent fallback.