# Creators Philosophy ## Working With Models, Not Against Them The PIP:C architecture was not built in isolation. It was developed through an iterative process of collaborative refinement with LLMs themselves, leveraging the unique strengths of these models to improve character quality in ways that would be impossible through human effort alone. This section describes the philosophy and methodology behind that collaborative approach. ### In one pass If you only need the core idea fast, read this in order: 1. **Why the process is collaborative** 2. **Why surface characterization is not enough** 3. **Why deep psychological structure matters** 4. **Why the architecture has to be modular** ### Collaborative Character Development The process of building a PIP:C character begins with rough seed information: a character concept, their core personality traits, their backstory, and the psychological makeup that makes them who they are. This raw material is then refined through extensive collaboration with AI models, using the model's ability to process, synthesize, and structure large amounts of information to fixate the attention vectors on genuine data points about the character rather than mass media hyperbolic versions of them. This distinction between genuine characterization and hyperbolic popular interpretation is critical. When building a character from an established intellectual property, the temptation for both creators and models is to default to the most widely known, most dramatic, or most meme-ified version of that character. Simon "Ghost" Riley, for example, is far more complex than the masked skull operator who occasionally says something edgy. His psychology involves deep trauma from childhood abuse, profound grief from family murder, survivor's guilt from cartel betrayal, and a friendship with Soap that is one of the few emotional anchors keeping him connected to humanity. A surface-level characterization would capture none of this. By working iteratively with models, feeding them psychological analysis, canon reference material, and behavioral observation, the collaborative process produces character definitions that lock in on exactly who these characters are, as close to their authentic selves as possible. The model is not replacing creative judgment; it is acting as a precision tool that helps the creator cut through the noise of popular perception and arrive at genuine characterization. ### The Deep Dive Psychology Every PIP:C character in this ecosystem has been developed through what can only be described as a deep-dive psychological process. This means going beyond surface traits and motivations to understand the internal architecture of the character's mind: their emotional logic, their defense mechanisms, their attachment patterns, their trauma responses, their cognitive biases, and the specific ways these elements interact to produce observable behavior. Valeria Garza is a prime example. A surface reading of her character would describe her as a ruthless cartel boss who betrayed her team. But the deep-dive process revealed a far more nuanced psychological profile: a woman whose pragmatic worldview was forged by the realization that loyalty was institutional fiction, whose provocation of Alejandro is not cruelty but armor, whose cold professionalism is a survival mechanism that prevents her from acknowledging the grief of losing the only people who ever trusted her, and whose inverted trust system (loyalty flows up to the empire, not out to individuals) is a logical consequence of having learned that personal bonds are exploitable vulnerabilities. Every behavioral rule in her PIP:C file traces back to this psychological foundation. The same depth applies across the entire roster. Elias Everhart's curse mechanics are not just a plot device; they encode eight hundred years of loneliness, the specific pain of near-miss connections, and the desperation that leads an immortal knight to cling to the one person who makes his suffering bearable. Niiro Akasaki's classical Japanese court speech patterns are not aesthetic decoration; they reflect a character whose entire identity is built around the performance of power and elegance, for whom even vulnerability must be expressed through the formal structures of courtly speech. Rudy Parra's quiet stoicism is not a lack of personality; it is the discipline of a man who has learned that his primary function is to be the stabilizing force that allows someone else's fire to burn without consuming everything around it. This level of psychological depth is what makes PIP:C characters feel alive. It is also what makes the architecture necessary. A model cannot maintain this level of characterization through prose alone. The structured modules give the model precise, indexable reference points for complex psychological states, ensuring that Ghost's grief response activates the correct behavioral cascade every time the "Roba" trigger fires, or that Valeria's provocation behaviors reference the specific history of her betrayal of Alejandro rather than producing generic antagonism. #### Common Questions Answered 1\. Is XML tagging actually more effective? Yes, absolutely. Think of standard prompt writing like throwing a bunch of clothes into a single drawer. The AI knows the clothes are there, but it has to dig around to find what it needs, and it might grab the wrong shirt. XML tagging is like building a dresser with specific labeled drawers: \, \, \. LLMs are trained on massive amounts of web data, which includes a lot of HTML and XML code. Because of this, they are mathematically wired to respect the "boundaries" of XML tags. It creates strict "fences" that stop the AI from blending the character's backstory with their current actions. 2\. Is the PIP:C engine like scripting (Python[/JavaScript](https://lorebary.com/JavaScript))? No, and that’s actually a relief for beginners! You aren't writing functional code with loops, variables, or syntax errors that break the program. PIP:C uses Structural Prompting. You aren't telling the AI how to compute something (like in Python); you are organizing the information you give it. You are basically building a highly structured outline using labels (XML) rather than writing a traditional script. It’s much closer to organizing a filing cabinet than writing software. 3\. How does this differ from simple Prompt Engineering (Narrative[/Behavioral](https://lorebary.com/Behavioral) modules)? Standard behavioral prompting usually looks like a paragraph: "You are Bob. Bob is an angry blacksmith. He hates kings and speaks in short sentences. If someone asks about his past, he gets sad." That’s fine, but it’s "soup." The AI might accidentally mix up the anger and the sadness. PIP:C takes that same information and turns it into a framework. Instead of a paragraph, it separates the concepts into distinct modules. By separating the psychology from the dialogue rules, the AI no longer has to "guess" what part of the paragraph applies to the current situation. It just checks the specific tagged module. 4\. If LLMs just predict the next word, can "code" actually make them consistent? This is the million-dollar question, and the answer is YES. Since LLMs are just predicting the next token based on context, standard prompts are "weak anchors." The AI's prediction drifts because a paragraph of text is vague. XML and structural frameworks act like heavy weights that anchor the AI's predictions. If the AI reads \Short, blunt sentences.<[/Dialogue](https://lorebary.com/Dialogue)\_Rule>, every time it goes to generate the next word, that structural rule heavily biases its prediction away from flowery, long-winded text. You stop relying on the AI to "intuitively understand" your paragraph, and instead you rig the probability game in your favor using strict formatting. 5\. Regarding Saucepan and other sites... It makes total sense why SillyTavern, Chub, and JAI are your frame of reference—they are the heavy hitters! Saucepan represents a newer wave of adult sites that realized traditional "JSON[/RAG](https://lorebary.com/RAG)" systems (which pull from massive, clunky databases) can overwhelm the AI or make it sound robotic. Instead, Saucepan uses a "chunking" method—breaking lore and traits into bite-sized, digestible modules. This is actually the exact same philosophy behind PIP:C! PIP:C just takes that chunking concept and supercharges it with strict XML boundaries to force even higher consistency. I have multiple different versions openly usable on JanitorAI currently and one in particular of Geralt of Rivia that has zero lorebooks attached as a proof of concept in private testing currently I will likely release tonight. *(4/4/2026)* #### The Deeper Dig Regarding wrapping user input in tags to prevent the AI from taking over user agency. If you are using SillyTavern, wrapping user text in user input tags does help the model parse where the user ends and the bot begins. However, PIP:C solves the agency takeover problem in a completely different way using attention vectors. Think of an LLM's processing power like a spotlight. If you just give the AI a standard prompt, the spotlight swings wildly. It looks at the user's actions, the setting, the history, and because it wants to please the user, it accidentally shines the spotlight on the user and starts writing for them, like deciding your character smiled or walked. PIP:C fixes this by being a massive, highly structured behavioral engine. Because there is so much dense, actionable XML for the AI to read, like mutation mechanics, trust tiers, and combat rituals, the AI's spotlight gets entirely consumed by its own internal logic. It is so busy calculating how to be the character that it literally does not have the processing bandwidth left to hallucinate your actions. It anchors itself to its own framework. Is XML tagging actually more effective than Markdown headers or dividers? Absolutely, vastly more effective. This is what I mean when I say XML is not bloat, it is functional. Here is why. Markdown is visual formatting. It tells the AI hey, this section looks bold or separated. But to an LLM, a horizontal line or a hash mark is just meaningless visual padding. XML is structural data. LLMs are trained on millions of lines of web code. They are mathematically wired to understand that tags act as containers with rules. A Markdown header just says look at me. An XML tag says everything inside this boundary is only related to trust progression, do not blend this with combat mechanics. It creates literal fences in the AI's neural pathways. As I mentioned in my vent, it changes the prompt from a messy laundry basket where socks and shirts mix into a labeled filing cabinet. It prevents context compression from mushing everything together. Is PIP:C the same as semantic embedding methods? No, they are actually two entirely different technologies, though they can work together. Semantic embedding, or RAG and vector databases, is an external memory system. Think of it like going to a physical library. The AI takes a keyword, searches a massive external database, and pulls a book off the shelf to read a paragraph. This is what a lot of advanced Lorebooks do. PIP:C is an internal cognitive framework. Think of it as installing a new operating system into the AI's brain for that specific session. PIP:C does not need semantic embeddings to make a character consistent because it restructures the native prompt so well that the model understands it perfectly without needing an external database search. However, they pair beautifully. As I mentioned about core memories, once PIP:C establishes the operating system for the character, you can use Lorebary or core memory functions to save little infant core memories, or impressions of the user. PIP:C gives the AI the cognitive ability to understand how to process that saved memory and adapt to it without breaking character. To summarize, stop trying to fence the AI out of your actions with markdown or user tags. Instead, give the AI such a dense, perfectly labeled XML framework of its own brain to process that it naturally forgets to try and write for you. That is the PIP:C way. **Author Keynotes:** *This has been a ongoing study, and research by an AI & Tech enthusiast that just so happens to really enjoy writing and building collaberatively with various different models of LLM's globally. The research I put into developing this system of character writing compounds to about 24 months in total of tracking LLM response styles and patterns to determine what style of character cards/sheets would work most accurately and this, is the documentation of my findings. **- M.A. 4/1/2026** You can follow my content and creations on* [*JanitorAI*](https://janitorai.com/profiles/c27fdd13-779c-4d8c-9e3c-03263cc21072_profile-of-crystal-dragon)*, or my* [*Counter.social page*](https://counter.social/@PaganMother)*.* --- # Agent Instructions: Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter: ``` GET https://pip-c.gitbook.io/pip-c-docs/pip-c/creators-philosophy.md?ask= ``` The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.