I would really like to see the character limit for ChatGPT Project Instructions substantially increased, or alternatively have an “extended/pinned instructions” system for complex Projects.

This isn’t just about wanting to write a bigger prompt. The current limit creates a structural problem for Projects that depend on persistent rules, especially long-form fiction, roleplaying games, worldbuilding, simulations, and D&D-style campaigns.

A campaign has two very different kinds of information:

1. Rules that should always apply

  • How the AI acts as GM/narrator

  • Player agency rules

  • Dice/check mechanics

  • POV and narration style

  • Who is allowed to control which characters

  • Hidden-information and spoiler rules

  • Continuity and retcon policies

  • Combat/stat/inventory procedures

  • Formatting and pacing requirements

  • How NPC knowledge is separated from narrator knowledge

2. Campaign state that changes over time

The first category belongs in Project Instructions.

The second belongs in Project memory, chats, and reference documents.

And importantly, OpenAI’s own documentation already recognizes this distinction. Its documentation for custom GPTs explicitly recommends using uploaded Knowledge for reference material and putting rules, tone, and workflow behavior in Instructions instead.

That is exactly the problem the current Project Instructions limit creates: once a sufficiently sophisticated rule set no longer fits, users are forced to put behavioral rules into documents, previous chats, or memory — the very places that should be holding the evolving state of the story.

That makes the system less reliable.

A rule such as “never decide the player character’s dialogue, thoughts, or voluntary actions” shouldn’t depend on whether the relevant campaign document happens to be referenced at that moment. Neither should combat procedures, POV restrictions, NPC knowledge rules, continuity procedures, or other foundational campaign mechanics.

Academic work on long-context LLMs supports this concern as well. Research such as Lost in the Middle (Liu et al., TACL 2024) has shown that models do not use all information in large contexts equally reliably. Information being somewhere in the available context is not the same thing as having a dedicated persistent instruction.

There is also an awkward memory problem.

Project-only Memory deliberately does not reference a user’s previously saved memories. That is a useful isolation feature, but it means global Memory is not an appropriate workaround for project-specific rules either. And putting campaign rules into project conversations means static behavioral instructions become mixed together with the dynamic history the Project is supposed to remember.

In other words:

The instruction limit forces us to use memory to remember how the AI should behave instead of using memory to remember what actually happened.

That matters enormously for a 50-session RPG campaign or a 200,000-word story.

OpenAI has already acknowledged the value of additional instruction space: in July 2026, Custom Instructions for paid users were increased from 1,500 to 5,000 characters specifically to provide more room for customizing ChatGPT’s behavior.

I think Projects need the same philosophy, arguably even more strongly, because Projects are specifically designed for long-running, evolving work.

My suggestion

Raise the Project Instructions limit substantially — perhaps to 20,000–32,000 characters for paid users.

Even better, provide two project-level layers:

Project Instructions
Always-on behavioral rules and workflows.

Project Sources / Memory
Lore, documents, conversations, facts, and evolving project state.

Another solution would be allowing users to mark a text document as an “Always-on Instruction Source”, meaning its contents are treated as Project Instructions rather than ordinary retrievable reference material.

I don’t think instructions should be unlimited. Extremely large or contradictory prompts can create their own problems, and concise instructions are still good practice.

But users shouldn’t have to choose between removing important behavioral rules and hiding those rules inside reference material that serves a fundamentally different purpose.

For normal Projects, the current limit may be perfectly adequate.

For storytelling, RPG campaigns, simulations, complex creative workflows, and other persistent systems, it becomes one of the main limitations preventing Projects from reaching their potential.

Let Project Instructions hold the rules, so Project Memory can remember the world.

I would really like to see the character limit for ChatGPT Project Instructions substantially increased, or alternatively have an “extended/pinned instructions” system for complex Projects.

This isn’t just about wanting to write a bigger prompt. The current limit creates a structural problem for Projects that depend on persistent rules, especially long-form fiction, roleplaying games, worldbuilding, simulations, and D&D-style campaigns.

A campaign has two very different kinds of information:

1. Rules that should always apply

  • How the AI acts as GM/narrator

  • Player agency rules

  • Dice/check mechanics

  • POV and narration style

  • Who is allowed to control which characters

  • Hidden-information and spoiler rules

  • Continuity and retcon policies

  • Combat/stat/inventory procedures

  • Formatting and pacing requirements

  • How NPC knowledge is separated from narrator knowledge

2. Campaign state that changes over time

The first category belongs in Project Instructions.

The second belongs in Project memory, chats, and reference documents.

And importantly, OpenAI’s own documentation already recognizes this distinction. Its documentation for custom GPTs explicitly recommends using uploaded Knowledge for reference material and putting rules, tone, and workflow behavior in Instructions instead.

That is exactly the problem the current Project Instructions limit creates: once a sufficiently sophisticated rule set no longer fits, users are forced to put behavioral rules into documents, previous chats, or memory — the very places that should be holding the evolving state of the story.

That makes the system less reliable.

A rule such as “never decide the player character’s dialogue, thoughts, or voluntary actions” shouldn’t depend on whether the relevant campaign document happens to be referenced at that moment. Neither should combat procedures, POV restrictions, NPC knowledge rules, continuity procedures, or other foundational campaign mechanics.

Academic work on long-context LLMs supports this concern as well. Research such as Lost in the Middle (Liu et al., TACL 2024) has shown that models do not use all information in large contexts equally reliably. Information being somewhere in the available context is not the same thing as having a dedicated persistent instruction.

There is also an awkward memory problem.

Project-only Memory deliberately does not reference a user’s previously saved memories. That is a useful isolation feature, but it means global Memory is not an appropriate workaround for project-specific rules either. And putting campaign rules into project conversations means static behavioral instructions become mixed together with the dynamic history the Project is supposed to remember.

In other words:

The instruction limit forces us to use memory to remember how the AI should behave instead of using memory to remember what actually happened.

That matters enormously for a 50-session RPG campaign or a 200,000-word story.

OpenAI has already acknowledged the value of additional instruction space: in July 2026, Custom Instructions for paid users were increased from 1,500 to 5,000 characters specifically to provide more room for customizing ChatGPT’s behavior.

I think Projects need the same philosophy, arguably even more strongly, because Projects are specifically designed for long-running, evolving work.

My suggestion

Raise the Project Instructions limit substantially — perhaps to 20,000–32,000 characters for paid users.

Even better, provide two project-level layers:

Project Instructions
Always-on behavioral rules and workflows.

Project Sources / Memory
Lore, documents, conversations, facts, and evolving project state.

Another solution would be allowing users to mark a text document as an “Always-on Instruction Source”, meaning its contents are treated as Project Instructions rather than ordinary retrievable reference material.

I don’t think instructions should be unlimited. Extremely large or contradictory prompts can create their own problems, and concise instructions are still good practice.

But users shouldn’t have to choose between removing important behavioral rules and hiding those rules inside reference material that serves a fundamentally different purpose.

For normal Projects, the current limit may be perfectly adequate.

For storytelling, RPG campaigns, simulations, complex creative workflows, and other persistent systems, it becomes one of the main limitations preventing Projects from reaching their potential.

Let Project Instructions hold the rules, so Project Memory can remember the world.