TL;DR
Several feature requests have already appeared independently in this community: hierarchical Projects, editing or excluding parts of long conversations, stronger reasoning in regular Chat, and switching between Chat and Work within the same conversation.
I do not think these are unrelated requests.
They point to a broader architectural principle:
Structure, conversation history, model capability, and execution mode should be independently controllable dimensions of a long-term AI workspace.
ChatGPT is increasingly used not only for one-off questions, but for research, learning, programming, writing, creative work, and projects that evolve over weeks or months. Those workflows are often nonlinear: users branch into side topics, abandon assumptions, reorganize projects, discover that a problem is harder than expected, or move gradually from discussion into execution.
The product should be able to accommodate that naturally.
1. Structure: Hierarchical Projects
Projects are useful for long-running work, but a flat list becomes difficult to manage as the number and complexity of Projects grows.
Users are often forced to choose between:
-
putting many related workstreams into one large Project, which mixes chats, files, instructions, and context; or
-
creating many independent top-level Projects, producing a long and structurally uninformative list.
A more scalable model would be:
Project
└── Subproject
└── Chat / Work / Files
Subprojects could optionally inherit selected instructions, files, or shared context from their parent while maintaining their own local context.
This would make hierarchy useful not only for visual organization, but also for context management.
Ideally, users could control what is inherited, what remains isolated, and whether search applies to the current Subproject, its parent, or the entire workspace.
2. History: Editable Conversation Context
Long conversations are essentially append-only.
Suppose a conversation contains:
A
B
C
D
E
F
G
and D–E become irrelevant, incorrect, or obsolete.
The user may still want A–C and F–G. Branching from C does not solve this problem, due to potentially useful chat responses after C.
Users should be able to select a message or message range and choose actions such as:
-
Delete from conversation
-
Exclude from active context
The second option is especially important.
A research discussion may contain a hypothesis that was later rejected. A programming conversation may contain obsolete code. A creative project may contain an abandoned direction.
Those messages can remain useful as history without continuing to influence future model responses.
In other words:
Visible conversation history and active model context need not be the same thing.
If later messages depended on removed context, ChatGPT could simply warn the user instead of preventing the edit.
3. Capability: Decouple Model Intelligence from Interaction Mode
Chat and Work represent different interaction styles.
Chat is conversational and exploratory.
Work is agentic and execution-oriented.
But task complexity and task format are not the same dimension.
Sometimes I need the strongest reasoning capability available to me simply to discuss a difficult mathematical argument, debug a conceptual problem, understand a paper, examine a definition, or explore a research question.
There might be no files to generate, websites to operate, autonomous tasks to complete, or concrete deliverable results to produce.
Sometimes the outcome is simply:
“I understand the problem now.”
That does not make the task intellectually simple.
Ideally, model capability and interaction style would be separate controls:
Capability
├── Fast
├── Medium
├── High
└── Highest capability available under the user's plan
Execution
├── Chat
└── Work
This would allow combinations such as:
Chat + High reasoning
Chat + Highest-capability model
Work + High reasoning
Work + Highest-capability model
Users should not need to adopt an agentic workflow merely because they need more reasoning capability.
The central principle is:
Task complexity and task format are different dimensions.
A task can be extremely difficult while remaining entirely conversational.
Conversely, an agentic Work task can be operationally simple.
4. Execution: Allow Natural Chat ↔ Work Transitions
Intellectual work often takes on different forms in its progress.
A conversation may begin with:
Why does this theorem require compactness?
then become:
What can we prove without compactness?
then:
This might be a useful lemma.
and eventually:
Let’s formally develop the argument.
The user cannot know at the beginning whether the conversation will remain exploratory or eventually require agentic execution.
So Chat and Work should ideally be stages of one continuous workflow.
A normal Chat should be able to escalate into Work when tools or autonomous execution are actually needed, then return to conversational Chat afterward while preserving the same history, files, and context.
Users should not have to predict the final form of a task before they begin thinking about it.
One Architecture Instead of Four Isolated Features
I think these requests fit naturally into four independent layers:
Structure
→ Project / Subproject
History
→ Conversation / editable active context
Capability
→ Model / reasoning level
Execution
→ Conversational Chat / agentic Work
Currently, some of these dimensions are structurally coupled in the product.
For long-term intellectual work, separating them would give users much more control without requiring ChatGPT to guess how every workflow should be organized.
Structural demand due to similar recurring proposals in the community
Several parts of this proposal have already been made independently in this community:
-
Hierarchical Projects, Custom Ordering & Subprojects
-
Delete or Split a Section of a Long Chat
-
Delete or Exclude Individual Messages from Context
-
Optional Extra High Reasoning in Existing Regular Chats
-
GPT-6 Astra Access on Plus Shouldn’t Be Limited to Work and Codex
-
Allow switching between Chat and Work within the same conversation
I do not see these as unrelated requests.
The fact that these requests all point to essentially the separation of different workflow needs suggests an inherent structural demand:
ChatGPT is becoming a long-term knowledge workspace, and users increasingly need independent control over how their work is structured, preserved as context, reasoned about, and executed.
I would love to see the product evolve in this direction.