cortex-implementation

Start by analyzing the exported flow JSON to identify Smartons, Formatters, HTTP requests, tags, and stages, then configure global settings, design the Cortex node prompt, replace Formatters with code tools, and wire downstream connections in Flowbuilder before testing.

Write an async function main(params) that normalizes the input, looks it up in a mapping object, persists results with await setField using the exact field keyword, and returns a small serializable object. Use the built-in http() function for external API calls since fetch and axios are unavailable.

The Cortex node prompt is a single text field limited to 10,000 characters. Draft and count characters before delivering, and compress by removing redundancy between sections rather than removing rules.

Use the HTTP Request UI component for a single declarative API call configured without code. Use a Code Tool with the http() function when you need chained calls, search-and-decide logic, idempotency, or orchestration across multiple systems in one step.

System fields like first_name and last_name can auto-capture any two-word answer, such as a vehicle model, when marked as globally recognizable. Add an explicit prompt rule that product data is never a name, and report persistent cases as a platform bug.

Confirmed limitations include calendar integration supporting only one mailbox, no native GPS/location recognition, and inactivity recovery that adapts by funnel stage rather than by specific flow branch. Reconfirm status with the product team before promising these to clients.