What problem does it solve? Coordinating work between a local agent and ChatGPT web sessions is error-prone: conversations lose context, duplicate messages get sent after interruptions, and the web side cannot execute local tasks. This Skill defines a disciplined protocol where the ChatGPT web side acts as the reasoning brain while the local agent handles transport and execution over a persistent conversation_id. ## Core Features & Use Cases - Four operating modes: Nudge mode (advance existing web sessions with minimal prompts), Directed Loop (web designs, local executes, results posted back), Peer Collaboration (blind answers, structured debate with steelmanning and disagreement reports), and Read-Only (pull conversation history, project files, and memories without sending anything). - Project-anchored session governance: All conversations must be confirmed with the user and bound to a ChatGPT project via project_id before the first message is sent; new sessions require explicit user approval. - Failure recovery discipline: Rules for interrupted sends, dead generation detection, composer draft cleanup, and daemon health checks prevent duplicate messages and stalled loops. - Use Case: You ask the local agent to let the web-side ChatGPT design an experiment; the agent opens a dedicated conversation in the right project, executes each instruction locally, posts factual results back to the same conversation_id, and continues until the web side declares completion. ## Quick Start Ask the agent to use the chatgpt-web bridge to continue your existing ChatGPT web conversation about a project, letting the web side decide the next steps while the local side only relays and executes.
