guides

Cookbooks

Each recipe is a worked example with a runnable request in curl, Python, and TypeScript, the actual output we received when we ran it, and a note on when to trust the result. Recipes cover three models: Decider 1 for typed decisions with probabilities (/v1/systemone), Restyler 1 for rewriting AI-sounding prose (/v1/restyle), and Query Fanout 1 for predicting the web searches an AI assistant is likely to run (/v1/fanout). Outputs are shown unedited, including cases where a model got something wrong.

Decider 1

• Route a support email in one call · Support inbox triage
• Act only when the model is sure · Confidence-gated actions
• Flag a phishing email · LLM guardrails
• Drop retrieved passages that do not help · RAG passage filtering
• Hold a reply that promises money · LLM guardrails
• Pick an agent's first tool · Agent routing and skill selection
• Ask many questions in one call · Structured extraction

Restyler 1

• Restyle a draft, then check the numbers · Content editing

Query Fanout 1

• Turn a buyer's question into a content brief · Generative engine optimisation
• See which rivals an assistant pulls in · Competitive research