Introducing Smallest AI Agent Skills: Give Your Coding Agent Voice API Expertise
Harshita Jain
Most of the time, an AI coding agent feels like a superpower. You describe what you want, and code appears in seconds. So when you ask one to add text-to-speech to your app, you expect the same. It hands back a clean function, you run it, and you get a 404. You spend the next two hours assuming you misconfigured something, before realizing the agent confidently called an API route that no longer exists. The code was never going to work, and the agent had no way of knowing that. Understanding why means understanding something about how these agents learn.
The problem your coding agent has with voice APIs
When you ask Claude Code or Cursor to add text-to-speech to your application, it usually writes something that looks correct.
Then you run it and get a 404.
Why?
Because the endpoint it used was deprecated six months ago.
AI coding agents are trained on snapshots of the internet, and APIs evolve much faster than model training cycles. Smallest AI's unified /waves/v1/tts endpoint replaced older per-model routes. Model names changed. Recommended sample rates changed. Best practices changed.
An agent relying on stale training data often generates code that was never going to work. You end up spending your afternoon debugging generated code instead of building your product.
There's a better way.
Introducing Smallest AI Agent Skills
Today, we're open-sourcing the Smallest AI Agent Skills collection:
GitHub Repository: https://github.com/smallest-inc/skills
Agent Skills is an open standard originally developed by Anthropic and now supported by Claude Code, Cursor, GitHub Copilot, Gemini CLI, and dozens of other AI coding tools.
Skills provide coding agents with specialized, up-to-date knowledge for a specific domain. Each skill is packaged as a SKILL.md file. When your agent detects a relevant task, it automatically loads the corresponding skill into context.
The result:
No hallucinated endpoints
No deprecated model names
No outdated examples
No API guesswork
Just working code.
Install in one line
Install the complete skill collection:
Or install only the skills you need:
Works with Claude Code, Cursor, GitHub Copilot (VS Code), Gemini CLI, Kiro, and any agent that supports the Agent Skills standard.
Six skills, everything covered
setup-api-key
Get your API key configured and verified. Start here.
Use this skill to:
Configure authentication
Verify API connectivity
Validate credentials
Troubleshoot setup issues
Example prompt:
"Set up my Smallest AI API key and verify that I can successfully make a request."
The agent loads the setup skill and walks through authentication using the latest onboarding flow.
text-to-speech
Everything you need for Lightning v3.1 and Lightning v3.1 Pro (Text-to-Speech models).
Includes:
HTTP generation
SSE streaming
WebSocket streaming
Voice cloning
Voice/model compatibility rules
Supported output formats
Sample rate recommendations
Example prompt:
"Generate speech from 'Your order has been confirmed.' using Smallest AI. Use the Magnus voice, stream it as WAV at 24000 Hz, and save it to output.wav."
The agent loads the Text-to-Speech skill and generates code using the correct endpoint, parameters, authentication headers, and streaming configuration.
speech-to-text
Everything you need for Pulse and Pulse Pro (Speech-to-Text models).
Includes:
Pre-recorded audio transcription
HTTP uploads
Real-time WebSocket streaming
Word-level timestamps
Speaker diarization
Automatic language detection
Example prompt:
"Transcribe recording.mp3 using Smallest AI. Return word-level timestamps and identify speaker changes."
The agent loads the Speech-to-Text skill and generates the correct transcription workflow using the latest API patterns.
speech-to-speech
Build real-time voice experiences with Hydra (Speech-to-Speech model).
Includes:
WebSocket session management
Turn detection
Barge-in handling
Tool calling
End-to-end voice pipeline examples
Example prompt:
"Create a real-time customer support voice agent using Hydra that can interrupt itself when the user starts speaking and call external tools when needed."
The agent loads the Speech-to-Speech skill and generates a complete real-time voice pipeline with the correct session and turn-handling logic.
llm-electron
Build with Electron using an OpenAI-compatible interface.
Includes:
Chat completions
Streaming responses
Function calling
Migration guidance from OpenAI models
Example prompt:
"Replace my OpenAI GPT-4o call with Smallest AI's Electron model. Keep the same messages array and streaming setup, just swap the base URL and model name."
The agent loads the Electron skill and performs the migration using the latest model names and API configuration.
voice-agents
Build production voice agents on Atoms.
Includes:
Agent creation
Agent crews
Outbound calling
Campaign management
Knowledge base integration
Call analytics
Example prompt:
"Create an outbound calling campaign for appointment reminders and give me code to monitor call outcomes and analytics."
The agent loads the Voice Agents skill and generates the correct Atoms API integration.
Getting Started
Detailed installation instructions, skill documentation, and example prompts are available in the documentation.
The skills repository is fully open source. If you find an issue, want additional examples, or would like to contribute improvements, pull requests are welcome.
Build faster. Ship working integrations. Let your coding agent stay up to date.
Resources:
How do I add or install skills in Claude Code?
How do I use skills in Claude Code?
Can I create my own custom skills?
Why do AI coding agents generate broken or deprecated API code?
