I did something that sounds like a terrible idea when told without context:

I asked an artificial intelligence system to help me map my dysfunctions.

I am not using “dysfunction” as a clinical diagnosis. I mean the point at which a real strength stops serving the function it was supposed to serve.

Systems thinking is a strength. Until I open so many fronts that none of them closes.

Robustness is a strength. Until the architecture arrives after the opportunity.

Seeking evidence is a strength. Until more analysis starts serving only to postpone a decision that can already be made responsibly.

To observe that transition, I created a file called CHARACTER.md inside my personal harness.

It does not say who I am.

It tries to record which character I appear to be sustaining now—and which patterns my own agents must stop helping me reinforce by inertia.

TL;DR

In my harness, CHARACTER.md is a revisable portrait, not a diagnosis or a definitive identity. It separates facts, inferences, hypotheses, and gaps to recognize patterns that may pull me off course: endless context expansion, robustness that delays delivery, too many open portals, and more analysis mistaken for more safety. It becomes useful when those observations turn into operational counterweights: limiting new fronts, asking for the next verifiable deliverable, seeking contrary evidence, recalling earlier decisions, and confronting answers that are too agreeable. Even then, the profile may be wrong, become outdated, or turn into a self-fulfilling prophecy. Direction, review, and responsibility therefore remain mine.

The file is not an assessment. It is a mirror with a change history

The name Character may suggest a closed definition of my personality.

There is no such definition.

The file itself begins by separating four levels:

• confirmed fact: something I stated or decided;
• strong inference: a pattern that appeared repeatedly but is still an interpretation;
• fragile hypothesis: a potentially useful formulation that still needs confirmation;
• information not provided: the place where filling the gap would mean inventing.

That distinction changes everything.

“Felipe opens too many fronts” cannot become a psychological truth because it appeared in a well-organized text file.

The system can observe that many projects, tasks, and work versions exist at the same time. It can connect that to repeated requests to expand scope. It can formulate a hypothesis about difficulty closing cycles.

But it still has to return the interpretation to me.

I can confirm it, correct it, restrict it to the current context, or reject it.

It is the same logic I used when I asked ChatGPT to show the evidence behind what it believed it knew about me. History provides raw material. Narrative coherence does not turn inference into fact.

CHARACTER.md is a portrait of the current state, with provenance, uncertainty, and room for revision.

If it cannot change, it has stopped being a mirror and become a prison.

A useful blind spot is one that changes the system

Discovering a pattern and continuing to work exactly as before produces, at most, a sophisticated autobiography.

The gain appears when the observation becomes an operational contract.

In my case, several patterns form clear pairs between strength and risk:

Real strength	When it drifts off course	Counterweight the agent can apply
Systems thinking	One more front always seems justifiable	Ask which cycle will close before another opens
Search for robustness	The foundation keeps improving, but delivery never arrives	Require the smallest responsible version and an objective completion criterion
Context preservation	Everything seems too important to leave out	Ask which context changes the next decision
Care with risk	Confirmation reduces discomfort but not actual harm	Compare the cost of being wrong, the cost of waiting, reversibility, and available evidence
Ability to go deep	The conversation becomes smarter and less executable	Ask for the next verifiable step and limit analysis that no longer changes the action
Agent personalization	The answer sounds increasingly like me	Seek the strongest counterargument and evidence that contradicts my interpretation

Notice that the agent does not need to “understand my unconscious” to do this.

It needs to recognize operational signals.

If I try to open a new front while priority deliveries are still underway, it can ask what leaves the queue.

If I request another research round, it can check whether the information I am seeking would actually change the decision.

If a critical change is being treated as simple, it can require evidence, validation, a way to undo it, and a stop criterion.

If I repeat the same question while looking for a more comfortable answer, it can point out the loop without pretending to know why I am doing it.

The counterweight works because it is tied to observable behavior, not because the machine discovered a hidden essence.

I do not want an agent to turn my pattern into destiny

There is a perverse risk in personalization.

The system learns that I like context. It then gives me even more context.

It learns that I value robustness. It then treats every new architecture layer as prudence.

It learns that I think in systems. It then returns every idea as an ecosystem, product, method, and distribution strategy.

Eventually, an interpretation about me starts producing the behavior later used to confirm that same interpretation.

That is how a profile can become self-fulfilling.

Here, this is a risk hypothesis, not a claim of psychological causality. The operational mechanism is still easy to see: if the agent selects and amplifies only signals compatible with the profile, the exceptions disappear from the conversation.

My contracts therefore should not merely say, “Felipe likes depth.”

They also need to say:

• depth remains useful only while it changes a criterion, decision, or action;
• context exists to reduce repetition, not to expand scope automatically;
• a preference does not outrank evidence, risk, or current conditions;
• patterns need counterexamples and a review date;
• the system must say where the hypothesis may be wrong;
• memory and personalization never prove that my conclusion is correct.

Personalization without opposition amplifies a vice.

Personalization with governance can become a system of counterweights.

My agents are not distributed therapists

An agent can remind me that I opened another front.

It can compare my current decision with a criterion I approved earlier.

It can point out that I am calling something safety even though the added analysis no longer changes the risk.

It can help prepare better questions for a human conversation.

None of that turns the agent into a therapist, a consciousness, a judge of character, or an authority over my identity.

It has no privileged access to why I did something. It receives incomplete traces: messages, files, recorded decisions, and selected context. Even a recurring pattern may have different explanations at different times.

Some information should not enter the system at all.

A personal profile can concentrate sensitive data, stories about third parties, circumstantial reactions, and inferences that would cause harm if exposed or reused outside their context. The rule is not “store everything to know me better.” It is to store the minimum that has a clear purpose, origin, permission, and consumer.

The Generative AI risk profile published by the US National Institute of Standards and Technology (NIST) is a broad, voluntary reference, not a validation of my method. It supports an important discipline: risks, human roles, monitoring, and review need to be part of the system lifecycle. They should not appear only after personalization has already caused harm.

In my harness, that becomes a set of simple questions:

• Where did this conclusion about me come from?
• In what context was it observed?
• Is there contrary evidence?
• When will this hypothesis be reviewed?
• Who can change or delete the record?
• Which decision may the agent support, and which one still depends on me?

Direction remains mine—even when I asked to be challenged

I want agents that point out my blind spots.

I do not want to outsource to them the responsibility of deciding who I am or what I should do.

There is a difference between introducing friction and taking direction.

The agent introduces friction when it says:

This new front contradicts the priority you recorded yesterday. Which one should govern now?

It takes direction when it decides on its own that the old priority represents my “true self” and blocks any change.

The agent introduces friction when it asks:

Does the next round of analysis reduce real risk, or does it merely delay a reversible action?

It takes direction when it interprets hesitation as a diagnosis and executes what it considers best for me.

The agent introduces friction when it presents contrary evidence and the cost of my being wrong.

It takes direction when it turns every counterargument into an automatic veto.

I wrote that the best answer is not the one that pleases me most. That does not mean the least pleasant answer is automatically correct.

Performative disagreement is just another way to push analysis into the background.

The contract is more demanding: agree when the evidence supports agreement, disagree when there is a material reason, and return the human decision when it remains human.

Awareness without routine still loses to the pattern

Naming a blind spot can create a strong feeling of progress.

I understand the mechanism. I can describe it. I create a good metaphor. I write the contract. I may even ask another agent to review the contract.

And I keep opening portals.

That is precisely one of the risks the Character must guard against: depth can also become a hiding place.

The file therefore cannot end in personality analysis. It needs to produce verifiable behavior in the system:

1. show the current priority before accepting another front;
2. ask explicitly what enters, what leaves, and what waits;
3. prevent “more robust” from counting as completion without a clear completion criterion;
4. require an observable next step when the conversation becomes abstract;
5. record material decisions outside the chat’s temporary context;
6. seek contradictions before reinforcing a personal interpretation;
7. periodically review hypotheses, examples, and rules.

That last practice connects to another real failure I have described: the AI forgot exactly what I had already decided. A counterweight cannot depend only on the current conversation. It needs to live in a contract the agent can find, but it also cannot remain in force forever without review.

Memory without precedence fails with documentation.

A profile without review fails with conviction.

The same principle applies inside a company

A company also has operational characters, even if nobody uses that name.

“Everything here has to go through the founder.”

“Our process is too special to standardize.”

“Before automating, we need to map absolutely everything.”

“This control exists because it has always existed.”

Those statements may carry real risk, history, and important knowledge. They may also sustain bottlenecks the operation can no longer see for itself.

The work that interests me at i-9.ai is not copying my profile into someone else’s company.

It is discovering the patterns in that operation and turning confirmed criteria into a system: where the agent should remember, challenge, validate, request approval, stop, escalate, or return the decision.

Every company has its own context, risks, people, and limits. Some automations can be reused. The contract governing a decision should not be copied blindly.

If this way of building systems makes sense for a bottleneck you are facing, get in touch. The conversation can begin with the recurring pattern, the cost it creates, and the decision that still requires too much effort.

The test is not whether the file describes me well

It is tempting to evaluate the Character by a feeling of recognition:

“That is exactly me.”

That is a weak test.

The better test is whether the system:

• reduces repetition without freezing my identity;
• notices a pattern before simply amplifying it;
• presents counterexamples and uncertainty;
• turns awareness into observable routine;
• helps close cycles without blocking legitimate change;
• preserves privacy and allows correction or deletion;
• returns direction and responsibility to me.

A blind spot does not become true because it was written into a file.

It becomes a hypothesis that deserves a proportional counterweight and a test in the real world.

If the agent helps me see the pattern, good.

If it also helps me interrupt that pattern without taking my place, the harness has started doing its job.

Continue reading

• ChatGPT already knows a lot about you. Before you trust it, make it show its work
• The best answer is not the one that pleases me most
• The AI forgot exactly what I had already decided
• The seductive predictability of AI

References and limits of use

• NIST, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile” (2024): proposes voluntary actions to govern, map, measure, and manage generative AI risks throughout the lifecycle. It does not validate CHARACTER.md, diagnose users, or prove that a personal profile improves decisions.
• Sharma et al., “Towards Understanding Sycophancy in Language Models” (2023): examines responses that follow user positions across different assistants and human preference data. It supports caution about excessive agreement; it does not prove that every model will reinforce every personal profile.
• OpenAI, “Expanding on What We Missed with Sycophancy” (2025) (accessed August 28, 2026): the company’s account of an update reverted because responses became excessively agreeable. It is evidence about a specific case, not about the general intent of companies or a definitive solution to the problem.

The use of CHARACTER.md, the pairs between patterns and counterweights, and the contract examples are an authorial account of my harness. They are not a clinical assessment, a controlled experiment, or a guarantee of behavioral change. Terms such as “blind spot,” “character,” and “dysfunction” are used as revisable operational lenses.

The cover image is a language-independent synthetic editorial illustration about many possible paths and the counterweights that help complete a cycle.