Lectures

In this lecture, we discuss how to design multistep workflows in ChatGPT Apps: break the scenario into steps, distribute responsibility between the GPT, the widget, and MCP/ Agent, describe steps as a mini state-machine (types like GiftWorkflowStep and a getNextStep function), and improve UX with progress and small screens instead of a single “mega-form.” We’ll also look at a mermaid diagram, examples in React ( useState, useWidgetState), and common mistakes when implementing multistep scenarios.

We unpack the tool gating pattern: why “all tools at once” degrades the quality of tool-call selection, bloats the context, and creates security risks; how to design a scenario so that the model sees only the tools it needs at the current step ( INTERVIEW → BROWSING → CHECKOUT), and how to combine model-level gating (via the current state and the list of tools) with logical checks inside handlers (for example, disallow checkout when the cart is empty). We also tie tool gating to the UI/widget and the MCP architecture.

We break down how to design and store scenario (workflow) context in ChatGPT Apps: what distinguishes UI State from LLM Context and Business State, how to describe the WorkflowContext structure in TypeScript, where to store data (prompt, widgetState, backend/MCP), how to introduce a workflowId, and how to resume a scenario, support “back/forward,” and avoid desynchronization between the UI, the model, and the server.

In this lecture, we cover how to design fault‑tolerant workflows in ChatGPT Apps: classify failures (technical, model, business errors and UX interruptions), pick strategies ( retry, fail-fast, rollback, and involving the user), implement exponential backoff via callWithRetry, protect critical steps with an idempotencyKey, build compensations using the Saga pattern, synchronize UI/backend/model steps, and present errors so both the model (a structure with code and retryable) and the user (brief, safe messages) can understand them. We finish with hands‑on practice on the GiftGenius checkout step and a checklist of common pitfalls.

In this lecture we break down how to measure workflow effectiveness in the ChatGPT App at the step level: we build a scenario funnel, define basic metrics ( start rate, completion rate, conversion, drop-off), design an event schema ( "workflow_started", "step_completed", "step_failed"), instrument the widget on Next.js and the server/MCP, learn to read the data and run A/B experiments, without forgetting about privacy and data minimization.