Feature proposal: chatgpt core directives + a new behavioral feedback infrastructure

FEATURE PROPOSAL: CHATGPT CORE DIRECTIVES + A NEW BEHAVIORAL FEEDBACK INFRASTRUCTURE

A product concept developed from the perspective of heavy ChatGPT users

Keegan Kachowski

THE OPPORTUNITY

ChatGPT has accomplished something extraordinary: one general-purpose AI can serve an enormous number of people with completely different needs.

But that creates a fundamental limitation.

What makes ChatGPT better for one user can make it worse for another.

One user wants concise answers. Another wants detailed explanations.

One user wants the AI to challenge them. Another wants it to execute instructions directly.

One user wants a particular formatting style consistently maintained. Another finds that same behavior frustrating.

There is a practical ceiling to how far one universal behavioral experience can go when hundreds of millions of users can have completely different expectations.

I believe the next major step in ChatGPT personalization is to separate what ChatGPT remembers about a user from what that user expects ChatGPT to consistently do.

THE CORE DIRECTIVES LAYER

I propose a small, persistent, user-specific Core Directives Layer.

Conceptually:

SYSTEM & SAFETY REQUIREMENTS

USER CORE DIRECTIVES

CONVERSATION CONTEXT & MEMORY

MODEL RESPONSE

Core Directives would contain a limited number of high-value behavioral expectations defined by the individual user.

They could describe communication preferences, formatting requirements, workflow expectations, interaction preferences, or other persistent behaviors.

The user defines WHAT matters.

The model retains flexibility in determining HOW to accomplish it.

This would not replace ChatGPT’s existing intelligence, memory, or safety architecture.

It would provide the user with a persistent steering mechanism for the part of the experience that is inherently personal.

The objective is not to create a different AI for every person.

The objective is to allow the same underlying intelligence to provide a meaningfully different experience to each individual user.

THE FEEDBACK LOOP IS WHERE THIS BECOMES MUCH MORE THAN PERSONALIZATION

The Core Directives system creates an opportunity far beyond giving users more customization.

When ChatGPT violates an established Core Directive, the user could identify the response as a Directive Violation.

That creates a fundamentally different feedback signal from a generic thumbs-down.

A thumbs-down tells OpenAI:

Something about this response disappointed this user.

A Directive Violation could tell OpenAI:

This user established a specific behavioral expectation, the model failed to satisfy it, and the user identified that specific failure.

The system already has the directive.

It already has the response.

It already has the surrounding conversation.

The user supplies the final piece:

This behavior was inconsistent with an expectation I explicitly established.

WHY THIS FEEDBACK COULD BE EXTREMELY VALUABLE

OpenAI already has enormous amounts of interaction data.

The challenge isn’t simply obtaining more data.

The challenge is extracting meaningful, actionable information from that data.

Traditional feedback often asks a user to stop what they are doing and provide information about their experience.

A busy user may give a quick rating, provide a short explanation, or ignore the request entirely.

The proposed system creates a different incentive.

The user encounters a problem while actively using ChatGPT.

That user already has a personal reason to report the problem because fixing it directly improves their own experience.

They aren’t completing a survey for OpenAI.

They are identifying a failure that matters to them.

That distinction could produce feedback that is more specific, more contextual, and more directly connected to an actual product failure.

THE AUTOMOTIVE ANALOGY

Imagine BMW, GM, Ford, or another manufacturer selling millions of vehicles.

Now imagine every vehicle could identify a problem while it was happening and report:

- what happened

- when it happened

- the conditions under which it happened

- which system was involved

- whether the problem had occurred previously

- how frequently it was occurring

- and relevant diagnostic information

The manufacturer would no longer have to rely entirely on customers noticing a problem, understanding it, finding a mechanic, explaining the symptoms, and eventually reporting it.

The product itself would become part of the diagnostic system.

That is the opportunity I see for ChatGPT.

When a persistent behavioral failure occurs, the system is already present when it happens.

The conversation is there.

The directive is there.

The response is there.

The user identifies the failure.

Instead of simply receiving:

THUMBS DOWN

OpenAI could potentially receive something closer to:

EXPECTED BEHAVIOR → ACTUAL BEHAVIOR → CONTEXT → USER-IDENTIFIED VIOLATION

That is a much richer signal.

THE INFORMATION COMES LOOKING FOR OPENAI

This may be one of the most important aspects of the proposal.

Traditional feedback collection often requires organizations to go looking for information through surveys, interviews, user studies, feedback requests, and extensive analysis.

The proposed system changes the direction of that information flow.

The users are already using the product.

The users already care about their own experience.

When something violates an expectation important enough to them, they have a reason to report it.

The feedback doesn’t necessarily have to be hunted down.

The user’s own experience creates the reason to provide it.

This could create a continuous source of highly contextualized behavioral feedback generated during real-world use.

FROM INDIVIDUAL FAILURE TO POPULATION-LEVEL INSIGHT

One user reports one violation.

That is useful for that user.

A thousand users report similar violations.

Now there may be a pattern.

A hundred thousand users report similar violations.

Now OpenAI may have evidence of a broader behavioral issue.

At scale, the platform could potentially develop a continuously generated dataset showing:

WHAT USERS EXPECT

WHERE MODELS FAIL

UNDER WHAT CONDITIONS THEY FAIL

HOW FREQUENTLY THEY FAIL

WHICH FAILURES ARE INDIVIDUAL

WHICH FAILURES MAY BE SYSTEMIC

This is fundamentally different from simply collecting satisfaction ratings.

And importantly, not every individual preference needs to become a global model change.

A behavior that works beautifully for one person may be completely wrong for another.

Core Directives remain individual.

Aggregated violation data can reveal broader patterns without forcing one user’s preferences onto everyone else.

ONE SYSTEM, TWO LEVELS OF VALUE

INDIVIDUAL LEVEL

ChatGPT becomes more consistent with the individual user’s expectations.

SYSTEM LEVEL

OpenAI receives structured information about recurring behavioral failures across its user population.

This creates a feedback architecture where personalization itself becomes an instrumentation layer for improving the product.

THE BUSINESS CASE

This is not simply a request for another personalization setting.

It potentially addresses several problems simultaneously:

- Less repetitive correction from users

- Greater long-term behavioral consistency

- Less prompt engineering required from users

- More meaningful personalization

- Higher-signal behavioral feedback

- Better visibility into recurring model failures

- Potentially less manual interpretation of ambiguous feedback

- A scalable mechanism for individualized AI experiences

And it does this without requiring OpenAI to create a completely different base model for every individual.

The shared intelligence remains shared.

The safety architecture remains authoritative.

The individual experience becomes customizable where it needs to be.

THE BIGGER OPPORTUNITY

OpenAI has spent years solving the enormous problem of building one AI capable of serving everyone.

I believe there is a natural ceiling to that approach.

You cannot make one behavioral experience perfectly satisfy hundreds of millions of different people because those people genuinely want different things.

The answer may not be to make the universal model increasingly complicated in an attempt to satisfy every possible preference.

The answer may be to make the universal foundation extremely capable and then give each individual user control over the final layer of their experience.

DON’T REPLACE THE ONE-SIZE-FITS-ALL AI.

LET THE USER CUSTOMIZE THE FINAL FIT.

That final portion of the experience may be where the greatest difference between an ordinary AI assistant and a truly personal AI relationship is created.

THE PROPOSAL IN ONE SENTENCE

Give every user a persistent, user-specific behavioral layer, allow the model flexibility in how it satisfies those objectives, and turn explicit violations of those objectives into structured, high-context feedback that can improve both the individual experience and the product as a whole.

The result isn’t simply a more personalized ChatGPT.

It is a product that could become better at identifying its own behavioral failures through the people actually using it.

THE LARGER VISION

Don’t just build an AI that users can give feedback to.

Build an AI whose real-world use helps reveal exactly where it needs to improve.

The product becomes more personal for the individual user while simultaneously becoming a richer source of information for the organization building it.

That creates a potentially powerful cycle:

USER EXPECTATION

MODEL BEHAVIOR

VIOLATION IDENTIFIED

STRUCTURED FEEDBACK

PRODUCT INSIGHT

IMPROVED SYSTEM

BETTER USER EXPERIENCE

INVITATION TO DISCUSS

This proposal is being presented from the perspective of a heavy ChatGPT user who has spent substantial time observing the practical limitations of long-term behavioral consistency, personalization, memory, and user feedback.

I am not presenting this as a claim that I have solved every technical implementation detail.

I am presenting it as a product concept that I believe is worth serious technical and commercial consideration.

If anyone at OpenAI, or anyone working in AI research, product development, personalization, human-AI interaction, model evaluation, or related fields is interested in discussing the proposal further, I would be very interested in continuing the conversation.

I would be happy to provide the complete proposal, explain the reasoning behind the concept, discuss possible implementation approaches, answer technical or product questions, and explore how the idea could potentially be adapted or improved.

For inquiries keegansproposal at gmail dot com

FINAL THOUGHT

The goal is not to tell OpenAI how to build its models.

The goal is to identify a problem from the user’s side and propose a mechanism that could potentially solve several problems at once.

ChatGPT has already reached a point where one AI can serve an extraordinary number of different people.

The next challenge is making that same intelligence feel genuinely personal without sacrificing the advantages of a shared platform.

We believe the answer may be found in giving the user control over the final layer of the experience while turning that personalization layer into a source of highly contextualized product feedback.

One platform.

One underlying intelligence.

Millions of individual experiences.

And a feedback loop that gets stronger as the platform grows.

Keegan Kachowski