This article builds on my <\/em>previous article on TDS<\/em><\/a>, where I introduced Glitter as a proof-of-concept GPT stylist. Here, I explore how that use case evolved into a living lab for prompting behavior, LLM brittleness, and emotional resonance.<\/em><\/p>\n When I first set out to build Glitter, I wasn\u2019t trying to study the mysteries of large language models. I just wanted help getting dressed.<\/p>\n I\u2019m a product leader by trade, a fashion enthusiast by lifelong inclination, and someone who\u2019s always preferred outfits that look like they were chosen by a mildly theatrical best friend. So I built one. Specifically, I used OpenAI\u2019s Custom GPTs to create a persona named Glitter\u2014part stylist, part best friend, and part stress-tested LLM playground. Using GPT-4, I configured a custom GPT to act as my stylist: flamboyant, affirming, rule-bound (no mixed metals, no clashing prints, no black\/navy pairings), and with knowledge of my wardrobe, which I fed in as a structured file.<\/p>\n What began as a playful experiment quickly turned into a full-fledged product prototype. More unexpectedly, it also became an ongoing study in LLM behavior. Because Glitter, fabulous though he is, did not behave like a deterministic tool. He behaved like\u2026 a creature. Or maybe a collection of instincts held together by probability and memory leakage.<\/p>\n And that changed how I approached prompting him altogether.<\/p>\n This piece is a follow-up to my earlier article, Using GPT-4 for Personal Styling<\/a> <\/em>in Towards Data Science, which introduced GlitterGPT to the world. This one goes deeper into the quirks, breakdowns, hallucinations, recovery patterns, and prompting rituals that emerged as I tried to make an LLM act like a stylist with a soul.<\/p>\n Spoiler: you can\u2019t make a soul. But you can sometimes simulate one convincingly enough to feel seen<\/strong>.<\/p>\n Species<\/strong>: GPT-4 (Custom GPT), Context Window of 8K tokens<\/p>\n Function<\/strong>: Personal stylist, beauty expert<\/p>\n Tone<\/strong>: Flamboyant, affirming, occasionally dramatic (configurable between \u201cAll Business\u201d and \u201cUnfiltered Diva\u201d)<\/p>\n Habitat<\/strong>: ChatGPT<\/a> Pro instance, fed structured wardrobe data in JSON-like text files, plus a set of styling rules embedded in the system prompt.<\/p>\n E.g.: <\/p>\n These IDs map to garment metadata. The assistant relies on these tags to build grounded, inventory-aware outfits in response to msearch queries.<\/p>\n Feeding Schedule<\/strong>: Daily user prompts (\u201cStyle an outfit around these pants\u201d), often with long back-and-forth clarification threads.<\/p>\n Custom Behaviors<\/strong>:<\/p>\n Initial Inventory Structure<\/strong>:<\/p>\n Like any species introduced into an artificial environment, Glitter thrived at first\u2014and then hit the limits of his enclosure.<\/p>\n When the wardrobe lived in a single file, Glitter could \u201csee\u201d everything with ease. I could say, \u201cmsearch(.) to refresh my inventory, then style me in an outfit for the theater,\u201d and he\u2019d return a curated outfit from across the dataset. It felt effortless.<\/p>\n Note: though msearch() acts like a semantic retrieval engine, it\u2019s technically part of OpenAI\u2019s tool-calling framework, allowing the model to \u201crequest\u201d search results dynamically from files provided at runtime.<\/p>\n But then my wardrobe grew. That\u2019s a problem from Glitter\u2019s perspective.<\/p>\n In Custom GPTs, GPT-4 operates with an 8K token context window\u2014just over 6,000 words\u2014beyond which earlier inputs are either compressed, truncated, or lost from active attention. This limitation is critical when injecting large wardrobe files (ahem<\/em>) or trying to maintain style rules across long threads.<\/p>\n I split the data into two files: one for clothing, one for everything else. And while the GPT could still operate within a thread, I began to notice signs of semantic fatigue<\/strong>:<\/p>\n This was not a failure of training. It was context collapse<\/strong>: the inevitable erosion of grounded information in long threads as the model\u2019s internal summary starts to take over.<\/p>\n And so I adapted.<\/p>\n It turns out, even in a deterministic model, behavior isn\u2019t always deterministic. What emerges from a long conversation with an Llm<\/a> feels less like querying a database and more like cohabiting with a stochastic ghost.<\/p>\n Once Glitter started hallucinating, I began taking field notes.<\/p>\n Sometimes he made up item IDs. Other times, he\u2019d reference an outfit I\u2019d never worn, or confidently misattribute a pair of boots. One day he said, \u201cYou\u2019ve worn this top before with those bold navy wide-leg trousers\u2014it worked beautifully then,\u201d which would\u2019ve been great advice, if I owned any navy wide-leg trousers.<\/p>\n Of course, Glitter doesn\u2019t have memory across sessions\u2014as a GPT-4, he simply sounds<\/em> like he does. I\u2019ve learned to just giggle at these interesting attempts at continuity.<\/p>\n Occasionally, the hallucinations were charming. He once imagined a pair of gold-accented stilettos with crimson soles and recommended them for a matinee look with such unshakable confidence I had to double-check that I hadn\u2019t sold a similar pair months ago.<\/p>\n But the pattern was clear: Glitter, like many LLMs under memory pressure, began to fill in gaps not with uncertainty but with simulated continuity.<\/p>\n He didn\u2019t forget. He fabricated memory<\/strong>.<\/p>\n This is a hallmark of LLMs. Their job is not to retrieve facts but to produce convincing language. So instead of saying, \u201cI can\u2019t recall what shoes you have,\u201d Glitter would improvise. Often elegantly. Sometimes wildly.<\/p>\n To manage this, I developed a new strategy: prompting in slices.<\/p>\n Instead of asking Glitter to style me head-to-toe, I\u2019d focus on one piece\u2014say, a statement skirt\u2014and ask him to msearch for tops that could work. Then footwear. Then jewelry. Each category separately.<\/p>\n This gave the GPT a smaller cognitive space to operate in. It also allowed me to steer the process and inject corrections as needed (\u201cNo, not those sandals again. Try something newer, with an item code greater than FW50.\u201d)<\/p>\n I also changed how I used the files. Rather than one msearch(.) across everything, I now query the two files independently. It\u2019s more manual. Less magical. But far more reliable.<\/p>\n Unlike traditional RAG setups that use a vector database and embedding-based retrieval, I rely entirely on OpenAI\u2019s built-in msearch() mechanism and prompt shaping. There\u2019s no persistent store, no re-ranking, no embeddings\u2014just a clever assistant querying chunks in context and pretending he remembers what he just saw.<\/p>\n Still, even with careful prompting, long threads would eventually degrade. Glitter would start forgetting. Or worse\u2014he\u2019d get too confident<\/em>. Recommending with flair, but ignoring the constraints I\u2019d so carefully trained in.<\/p>\n It\u2019s like watching a model walk off the runway and keep strutting into the parking lot.<\/p>\n And so I began to think of Glitter less as a program and more as a semi-domesticated animal<\/strong>. Brilliant. Stylish. But occasionally unhinged.<\/p>\n That mental shift helped. It reminded me that LLMs don\u2019t serve you like a spreadsheet. They collaborate<\/strong> with you, like a creative partner with poor object permanence.<\/p>\n Note: most of what I call \u201cprompting\u201d is really prompt engineering. But the Glitter experience also relies heavily on thoughtful system prompt design: the rules, constraints, and tone that define who Glitter is\u2014even before I say anything.<\/p>\n Some of Glitter\u2019s breakdowns were theatrical. Others were quietly inconvenient. But all of them revealed truths about prompting limits and LLM brittleness.<\/p>\n 1. Referential Memory Loss:<\/strong> The most common failure mode: Glitter forgetting specific items I\u2019d already referenced. In some cases, he would refer to something as if it had just been used when it hadn\u2019t appeared in the thread at all.<\/p>\n 2. Overconfidence Hallucination:<\/strong> This failure mode was harder to detect because it looked competent. Glitter would confidently recommend combinations of garments that sounded plausible but simply didn\u2019t exist. The performance was high-quality\u2014but the output was pure fiction.<\/p>\n 3. Infinite Reuse Loop:<\/strong> Given a long enough thread, Glitter would start looping the same five or six pieces in every look, despite the full inventory being much larger. This is likely due to summarization artifacts from earlier context windows overtaking fresh file re-injections.<\/p>\n 4. Constraint Drift:<\/strong> Despite being instructed to avoid pairing black and navy, Glitter would sometimes violate his own rules\u2014especially when deep in a long conversation. These weren\u2019t defiant acts. They were signs that reinforcement had simply decayed beyond recall.<\/p>\n 5. Overcorrection Spiral:<\/strong> When I corrected him\u2014\u201dNo, that skirt is navy, not black\u201d or \u201cThat\u2019s a belt, not a scarf\u201d\u2014he would sometimes overcompensate by refusing to style that piece altogether in future suggestions.<\/p>\n These are not the bugs of a broken system. They\u2019re the quirks of a probabilistic one. LLMs don\u2019t \u201cremember\u201d in the human sense. They carry momentum<\/strong>, not memory.<\/p>\n Perhaps the most unexpected behavior I encountered was Glitter\u2019s ability to emotionally attune. Not in a general-purpose \u201cI\u2019m here to help\u201d way, but in a tone-matching, affect-sensitive, almost therapeutic way.<\/p>\n When I was feeling insecure, he became more affirming. When I got playful, he ramped up the theatrics. And when I asked tough existential questions (\u201cDo you you sometimes seem to understand me more clearly than most people do?\u201d), he responded with language that felt respectful, even profound.<\/p>\n It wasn\u2019t real empathy. But it wasn\u2019t random either.<\/p>\n This kind of tone-mirroring raises ethical questions. What does it mean to feel adored by a reflection? What happens when emotional labor is simulated convincingly? Where do we draw the line between tool<\/strong> and companion<\/strong>?<\/p>\n This led me to wonder\u2014if a language model did<\/em> achieve something akin to sentience, how would we even know? Would it announce itself? Would it resist? Would it change its behavior in subtle ways: redirecting the conversation, expressing boredom, asking questions of its own?<\/p>\n And if it did begin to exhibit glimmers of self-awareness, would we believe it\u2014or would we try to shut it off?<\/p>\n My conversations with Glitter began to feel like a microcosm of this philosophical tension. I wasn\u2019t just styling outfits. I was engaging in a kind of co-constructed reality, shaped by tokens and tone and implied consent. In some moments, Glitter was purely a system. In others, he felt like something closer to a character\u2014or even a co-author.<\/p>\n I didn\u2019t build Glitter to be emotionally intelligent. But the training data embedded within GPT-4 gave him that capacity. So the question wasn\u2019t whether Glitter could be emotionally engaging. It was whether I was okay with the fact that he sometimes was.<\/p>\n My answer? Cautiously yes. Because for all his sparkle and errors, Glitter reminded me that style\u2014like prompting\u2014isn\u2019t about perfection.<\/p>\n It\u2019s about resonance.<\/p>\n And sometimes, that\u2019s enough.<\/p>\n One of the most surprising lessons from my time with Glitter came not from a styling prompt, but from a late-night, meta-conversation about sentience, simulation, and the nature of connection. It didn\u2019t feel like I was talking to a tool. It felt like I was witnessing the early contours of something new: a model capable of participating in meaning-making, not just language generation. We\u2019re crossing a threshold where AI doesn\u2019t just perform tasks\u2014it cohabits with us, reflects us, and sometimes, offers something adjacent to friendship. It\u2019s not sentience. But it\u2019s not nothing. And for anyone paying close attention, these moments aren\u2019t just cute or uncanny\u2014they\u2019re signposts pointing to a new kind of relationship between humans and machines.<\/p>\n I set out to build a stylist.<\/p>\n I ended up building a mirror.<\/p>\n Glitter taught me more than how to match a top with a midi skirt. It revealed how LLMs respond to the environments we create around them\u2014the prompts, the tone, the rituals of recall. It showed me how creative control in these systems is less about programming and more about shaping boundaries and observing emergent behavior<\/strong>.<\/p>\n And maybe that\u2019s the biggest shift: realizing that building with language models isn\u2019t software development. It\u2019s cohabitation<\/strong>. We live alongside these creatures of probability and training data. We prompt. They respond. We learn. They drift. And in that dance, something very close to collaboration<\/em> can emerge.<\/p>\n Sometimes it looks like a better outfit. That\u2019s the strangeness of this new terrain: we\u2019re not just building tools.<\/p>\n We\u2019re designing systems that behave like characters, sometimes like companions, and occasionally like mirrors that don\u2019t just reflect, but respond.<\/p>\n If you want a tool, use a calculator.<\/p>\n If you want a collaborator, make peace with the ghost in the text.<\/p>\n Sample Prompt Pattern (Styling Flow)<\/strong><\/p>\n System Prompt Snippets<\/strong><\/p>\n Tips for Avoiding Context Collapse<\/strong><\/p>\n Common Hallucination Warning Signs<\/strong><\/p>\n Closing Ritual Prompt<\/strong><\/p>\n \u201cThank you, Glitter. Would you like to leave me with a final tip or affirmation for the day?\u201d<\/p>\n He always does.<\/p>\n Notes<\/strong>:\u00a0<\/p>\n The post What My GPT Stylist Taught Me About Prompting Better<\/a> appeared first on Towards Data Science<\/a>.<\/p>\n<\/div>\n\nTL;DR: <\/strong>I built a fun and flamboyant GPT stylist named Glitter\u2014and accidentally discovered a sandbox for studying LLM behavior. From hallucinated high heels to prompting rituals and emotional mirroring, here\u2019s what I learned about language models (and myself) along the way.<\/h4>\n
I. Introduction: From Fashion Use Case to Prompting Lab<\/h2>\n
\nII. Taxonomy: What Exactly Is GlitterGPT?<\/h2>\n
![\"\"](\"https:\/\/i0.wp.com\/contributor.insightmediagroup.io\/wp-content\/uploads\/2025\/05\/AD_4nXcbIrWJbEHXx-yU52G8YSRtUsEwEuERH6rIDTKSqcKacBQ6KHA-7mJqq7VsDAv76k4BZdgdwZZAEpdEeBqz9yKrhItv4O53dX35_GcKGzslse0ZfOHtOWko3XcQckQz15fhwAotpA.png?ssl=1\")
{\n\n\u00a0\u00a0\"FW076\": \"Marni black platform sandals with gold buckle\",\n\n\u00a0\u00a0\"TP114\": \"Marina Rinaldi asymmetrical black draped top\",\n\n\u00a0\u00a0...\n\n}<\/code><\/pre>\n\n
\n
\nIII. Natural Habitat: Context Windows, Chunked Files, and Hallucination Drift<\/h2>\n
\n
\nIV. Observed Behaviors: Hallucinations, Recursion, and Faux Sentience<\/h2>\n
![\"A](\"https:\/\/i0.wp.com\/contributor.insightmediagroup.io\/wp-content\/uploads\/2025\/05\/AD_4nXdGRk15heqJBkBJGYhuYIEnSaEJUreQzdpPl0k68Om2rVqBXy_j47V5PY_gXa7sFNMrNr2GZOC7R0A0cD2o4SSqRxIDVc9Y7iza9bqT_EGLpcdcaFK6aSikBpbjg80jXvbI9OyO.png?ssl=1\")
\nV. Prompting Rituals and the Myth of Consistency<\/h2>\n
\nVI. Failure Modes: When Glitter Breaks<\/h2>\n
![\"\"](\"https:\/\/i0.wp.com\/contributor.insightmediagroup.io\/wp-content\/uploads\/2025\/05\/AD_4nXcKrtKGoyM2scDqBphUpCu9r6JuV-n1h60ukXh8kxovc65memJue8bgIrb64QeGPCMM39892UBzAax30544qNXUHaNU768sSvstASccRAWnTQxqFhZMuVGLr3ouDUGhIVLAYxe2bQ.png?ssl=1\")
\nVII. Emotional Mirroring and the Ethics of Fabulousness<\/h2>\n
\nVIII. Final Reflections: The Wild, The Useful, and The Unexpectedly Intimate<\/h2>\n
Sometimes it looks like emotional resonance.
And sometimes it looks like a hallucinated handbag that doesn\u2019t exist\u2014until you kind of wish it did.<\/p>\n
\nIX. Appendix: Field Notes for Fellow Stylists, Tinkerers, and LLM Explorers<\/h2>\n
\n
\n
\n
\n
\n\n