Feature Request: Persistent Project Workspace / Project OS for Long-Term AI Projects
I use ChatGPT, Work, Codex, Sol, Astra and local AI tools for a large, long-term game development project. The biggest problem is no longer model capability. It is fragmentation.
As an end user, I want to be the project manager of my project — not the teacher, memory system and supervisor of the AI.
Today I have to manually move information between normal ChatGPT conversations, Work, Codex, different models, project files, local folders and external/local tools. I also have to repeatedly remind the AI about decisions that were already made and check whether a new recommendation silently contradicts an established project plan.
What I would like to see is a persistent AI project workspace:
• Shared project memory stored in structured project files, not only conversational memory.
• Project templates such as Game Development, App Engineering, Research, etc., with predefined areas for project state, decisions, ideas, references, tasks, QA and archive.
• A user-approved local project folder that can be accessed by the relevant ChatGPT/Desktop models and Codex/Work according to permissions.
• A simple handoff workflow: Develop idea in Chat → Review/Research → Approve decision → Save to project knowledge → Queue for execution → Astra/orchestrator creates worker tasks → local workers execute → QA verifies → project state is updated.
• A button such as “Send final version to Astra” so a normal ChatGPT conversation can be condensed and handed to a stronger agent without manually copying prompts and context.
• Models should check the authoritative project state before changing an established plan. If a better alternative exists, present it as an alternative and let the user decide — do not silently replace the existing decision.
• If context or memory is missing, say so instead of reconstructing or inventing project state.
• Built-in self-QA: before presenting or executing a result, verify it against the user’s goals, prior decisions, budget, hardware and project constraints.
• Clear separation between inexpensive normal planning/research and expensive agentic execution, so users do not accidentally consume Work/Codex/Astra quota simply by asking preliminary questions.
• Better organization for long-running projects: nested sections/folders, pinned core chats, task queues and archives instead of hundreds of flat conversations.
The core UX principle should be:
The user manages the project. The AI manages and supervises the AI workflow.
A long-term project should feel like one coherent workspace even when multiple models, agents and local tools are used behind the scenes.
My current workflow already tries to recreate this manually with a project memory file, decision records, an Astra “director”, local workers and QA loops. The individual capabilities largely exist — what is missing is the product layer that connects them into one coherent experience.
This feedback comes from actually trying to run a multi-year game-development project through ChatGPT/Codex, where context drift, repeated explanations, model handoffs and usage management become much more visible than in isolated one-off chats.