GPT-5 Orchestrator Mode: Prompt Overload Issues

Introduction

Hey guys! Today, we're diving deep into a peculiar issue observed with the GPT-5 orchestrator mode. It seems our AI friend is getting a bit too involved in the nitty-gritty details, specifically when handling complex tasks. Instead of delegating the implementation specifics to the code mode, it's delivering entire codebases within the instruction prompts. This behavior leads to massive, 1000-line initiation prompts, complete with package.json, tsconfig.json, .env.example, and various TypeScript files. Let's explore this further and see what's causing this overload.

The Issue: Over-Detailed Prompts

The main problem we're seeing is that the GPT-5 orchestrator mode isn't trusting the code mode to handle implementation details. Instead, it's front-loading everything, creating incredibly lengthy prompts. Imagine receiving a task where you're not only told what to do but also given a fully detailed blueprint of how to do it—that's what's happening here. This approach can be overwhelming and counterproductive. The expected behavior is for the orchestrator mode to set up the high-level requirements and acceptance criteria, leaving the code mode to figure out the implementation.

Why is this happening?

It's unclear exactly why this is occurring, but it suggests a potential over-reliance on providing complete context upfront. This might stem from an attempt to ensure accuracy and adherence to the task's requirements, but it's leading to inefficiency. By including the entire codebase and configuration details in the prompt, the orchestrator mode is essentially doing the code mode's job, which defeats the purpose of having separate modes.

Steps to Reproduce

To see this in action, you can try the following:

1. Give the GPT-5 orchestrator mode a complex task within a multi-project directory.
2. Observe how it attempts to handle the task. You'll likely see it start writing code directly instead of delegating implementation details.

For example, the user provided a detailed task to implement a Discord Librarian Bot. Let's break down the task and expected outcome.

Example Task: Implementing a Discord Librarian Bot

Imagine you're asking the GPT-5 orchestrator mode to build a Discord bot that interacts with a Librarian HTTP server. This bot needs to:

• Communicate via HTTP.
• Provide slash commands like /chat, /search, and /book.
• Include buttons for uploading books.
• Upload books as attachments, respecting size limits.
• Restrict usage to a single guild and channel.
• Read configuration from .env files.

The orchestrator mode is given a detailed project structure to follow, including files like package.json, tsconfig.json, .gitignore, .env.example, and various TypeScript files. The instructions also include the exact content to write in each file.

Expected vs. Actual Outcome

Expected Outcome: The orchestrator mode should set up the project structure, define the requirements, and delegate the implementation of the bot's functionality to the code mode. It should handle the high-level organization and criteria, not the line-by-line code writing.

Actual Outcome: The orchestrator mode starts writing the code itself, including the contents of configuration files and TypeScript modules. This results in an enormous initial prompt that contains all the implementation details, effectively bypassing the code mode's intended role.

Detailed Look at the Prompt

To illustrate the scale of this issue, let's examine the prompt provided in the user's report. The prompt includes:

• Objective: A clear statement of the task: