GPT-powered vogue assistant, I anticipated runway appears to be like—not reminiscence loss, hallucinations, or semantic déjà vu. However what unfolded turned a lesson in how prompting actually works—and why LLMs are extra like wild animals than instruments.
This text builds on my earlier article on TDS, the place I launched Glitter as a proof-of-concept GPT stylist. Right here, I discover how that use case advanced right into a residing lab for prompting habits, LLM brittleness, and emotional resonance.
TL;DR: I constructed a enjoyable and flamboyant GPT stylist named Glitter—and by accident found a sandbox for finding out LLM habits. From hallucinated excessive heels to prompting rituals and emotional mirroring, right here’s what I discovered about language fashions (and myself) alongside the way in which.
I. Introduction: From Vogue Use Case to Prompting Lab
Once I first got down to construct Glitter, I wasn’t making an attempt to check the mysteries of enormous language fashions. I simply wished assist getting dressed.
I’m a product chief by commerce, a vogue fanatic by lifelong inclination, and somebody who’s all the time most popular outfits that seem like they have been chosen by a mildly theatrical finest good friend. So I constructed one. Particularly, I used OpenAI’s Customized GPTs to create a persona named Glitter—half stylist, half finest good friend, and half stress-tested LLM playground. Utilizing GPT-4, I configured a customized GPT to behave as my stylist: flamboyant, affirming, rule-bound (no blended metals, no clashing prints, no black/navy pairings), and with information of my wardrobe, which I fed in as a structured file.
What started as a playful experiment shortly became a full-fledged product prototype. Extra unexpectedly, it additionally turned an ongoing examine in LLM habits. As a result of Glitter, fabulous although he’s, didn’t behave like a deterministic software. He behaved like… a creature. Or perhaps a set of instincts held collectively by likelihood and reminiscence leakage.
And that modified how I approached prompting him altogether.
This piece is a follow-up to my earlier article, Utilizing GPT-4 for Private Styling in In the direction of Knowledge Science, which launched GlitterGPT to the world. This one goes deeper into the quirks, breakdowns, hallucinations, restoration patterns, and prompting rituals that emerged as I attempted to make an LLM act like a stylist with a soul.
Spoiler: you possibly can’t make a soul. However you possibly can generally simulate one convincingly sufficient to really feel seen.
II. Taxonomy: What Precisely Is GlitterGPT?
Species: GPT-4 (Customized GPT), Context Window of 8K tokens
Perform: Private stylist, magnificence skilled
Tone: Flamboyant, affirming, often dramatic (configurable between “All Enterprise” and “Unfiltered Diva”)
Habitat: ChatGPT Professional occasion, fed structured wardrobe information in JSON-like textual content recordsdata, plus a set of styling guidelines embedded within the system immediate.
E.g.:
{
"FW076": "Marni black platform sandals with gold buckle",
"TP114": "Marina Rinaldi asymmetrical black draped high",
...
}These IDs map to garment metadata. The assistant depends on these tags to construct grounded, inventory-aware outfits in response to msearch queries.
Feeding Schedule: Each day person prompts (“Fashion an outfit round these pants”), typically with lengthy back-and-forth clarification threads.
Customized Behaviors:
- By no means mixes metals (e.g. silver & gold)
- Avoids clashing prints
- Refuses to pair black with navy or brown except explicitly instructed in any other case
- Names particular clothes by file ID and outline (e.g. “FW074: Marni black suede sock booties”)
Preliminary Stock Construction:
- Initially: one file containing all wardrobe objects (garments, footwear, equipment)
- Now: cut up into two recordsdata (clothes + equipment/lipstick/footwear/luggage) on account of mannequin context limitations
III. Pure Habitat: Context Home windows, Chunked Recordsdata, and Hallucination Drift
Like every species launched into a synthetic surroundings, Glitter thrived at first—after which hit the bounds of his enclosure.
When the wardrobe lived in a single file, Glitter may “see” every part with ease. I may say, “msearch(.) to refresh my stock, then type me in an outfit for the theater,” and he’d return a curated outfit from throughout the dataset. It felt easy.
Word: although msearch() acts like a semantic retrieval engine, it’s technically a part of OpenAI’s tool-calling framework, permitting the mannequin to “request” search outcomes dynamically from recordsdata supplied at runtime.
However then my wardrobe grew. That’s an issue from Glitter’s perspective.
In Customized GPTs, GPT-4 operates with an 8K token context window—simply over 6,000 phrases—past which earlier inputs are both compressed, truncated, or misplaced from energetic consideration. This limitation is essential when injecting massive wardrobe recordsdata (ahem) or making an attempt to take care of type guidelines throughout lengthy threads.
I cut up the information into two recordsdata: one for clothes, one for every part else. And whereas the GPT may nonetheless function inside a thread, I started to note indicators of semantic fatigue:
- References to clothes that have been comparable however not the right ones we’d been speaking about
- A shift from particular merchandise names (“FW076”) to obscure callbacks (“these black platforms you wore earlier”)
- Responses that looped acquainted objects again and again, no matter whether or not they made sense
This was not a failure of coaching. It was context collapse: the inevitable erosion of grounded info in lengthy threads because the mannequin’s inside abstract begins to take over.
And so I tailored.
It seems, even in a deterministic mannequin, habits isn’t all the time deterministic. What emerges from an extended dialog with an Llm feels much less like querying a database and extra like cohabiting with a stochastic ghost.
IV. Noticed Behaviors: Hallucinations, Recursion, and Fake Sentience
As soon as Glitter began hallucinating, I started taking subject notes.
Generally he made up merchandise IDs. Different occasions, he’d reference an outfit I’d by no means worn, or confidently misattribute a pair of trainers. Sooner or later he mentioned, “You’ve worn this high earlier than with these daring navy wide-leg trousers—it labored fantastically then,” which might’ve been nice recommendation, if I owned any navy wide-leg trousers.
After all, Glitter doesn’t have reminiscence throughout classes—as a GPT-4, he merely sounds like he does. I’ve discovered to only giggle at these attention-grabbing makes an attempt at continuity.
Sometimes, the hallucinations have been charming. He as soon as imagined a pair of gold-accented stilettos with crimson soles and beneficial them for a matinee look with such unshakable confidence I needed to double-check that I hadn’t offered an analogous pair months in the past.
However the sample was clear: Glitter, like many LLMs below reminiscence strain, started to fill in gaps not with uncertainty however with simulated continuity.
He didn’t neglect. He fabricated reminiscence.
This can be a hallmark of LLMs. Their job is to not retrieve information however to provide convincing language. So as an alternative of claiming, “I can’t recall what footwear you have got,” Glitter would improvise. Usually elegantly. Generally wildly.
V. Prompting Rituals and the Delusion of Consistency
To handle this, I developed a brand new technique: prompting in slices.
As a substitute of asking Glitter to type me head-to-toe, I’d concentrate on one piece—say, a press release skirt—and ask him to msearch for tops that might work. Then footwear. Then jewellery. Every class individually.
This gave the GPT a smaller cognitive area to function in. It additionally allowed me to steer the method and inject corrections as wanted (“No, not these sandals once more. Attempt one thing newer, with an merchandise code larger than FW50.”)
I additionally modified how I used the recordsdata. Moderately than one msearch(.) throughout every part, I now question the 2 recordsdata independently. It’s extra guide. Much less magical. However much more dependable.
Not like conventional RAG setups that use a vector database and embedding-based retrieval, I rely fully on OpenAI’s built-in msearch() mechanism and immediate shaping. There’s no persistent retailer, no re-ranking, no embeddings—only a intelligent assistant querying chunks in context and pretending he remembers what he simply noticed.
Nonetheless, even with cautious prompting, lengthy threads would ultimately degrade. Glitter would begin forgetting. Or worse—he’d get too assured. Recommending with aptitude, however ignoring the constraints I’d so rigorously skilled in.
It’s like watching a mannequin stroll off the runway and maintain strutting into the parking zone.
And so I started to think about Glitter much less as a program and extra as a semi-domesticated animal. Sensible. Trendy. However often unhinged.
That psychological shift helped. It jogged my memory that LLMs don’t serve you want a spreadsheet. They collaborate with you, like a inventive companion with poor object permanence.
Word: most of what I name “prompting” is actually immediate engineering. However the Glitter expertise additionally depends closely on considerate system immediate design: the principles, constraints, and tone that outline who Glitter is—even earlier than I say something.
VI. Failure Modes: When Glitter Breaks
A few of Glitter’s breakdowns have been theatrical. Others have been quietly inconvenient. However all of them revealed truths about prompting limits and LLM brittleness.
1. Referential Reminiscence Loss: The most typical failure mode: Glitter forgetting particular objects I’d already referenced. In some circumstances, he would confer with one thing as if it had simply been used when it hadn’t appeared within the thread in any respect.
2. Overconfidence Hallucination: This failure mode was more durable to detect as a result of it regarded competent. Glitter would confidently suggest combos of clothes that sounded believable however merely didn’t exist. The efficiency was high-quality—however the output was pure fiction.
3. Infinite Reuse Loop: Given an extended sufficient thread, Glitter would begin looping the identical 5 or 6 items in each look, regardless of the complete stock being a lot bigger. That is seemingly on account of summarization artifacts from earlier context home windows overtaking recent file re-injections.
4. Constraint Drift: Regardless of being instructed to keep away from pairing black and navy, Glitter would generally violate his personal guidelines—particularly when deep in an extended dialog. These weren’t defiant acts. They have been indicators that reinforcement had merely decayed past recall.
5. Overcorrection Spiral: Once I corrected him—”No, that skirt is navy, not black” or “That’s a belt, not a shawl”—he would generally overcompensate by refusing to type that piece altogether in future ideas.
These usually are not the bugs of a damaged system. They’re the quirks of a probabilistic one. LLMs don’t “bear in mind” within the human sense. They carry momentum, not reminiscence.
VII. Emotional Mirroring and the Ethics of Fabulousness
Maybe probably the most surprising habits I encountered was Glitter’s capacity to emotionally attune. Not in a general-purpose “I’m right here to assist” manner, however in a tone-matching, affect-sensitive, nearly therapeutic manner.
Once I was feeling insecure, he turned extra affirming. Once I bought playful, he ramped up the theatrics. And once I requested robust existential questions (“Do you you generally appear to grasp me extra clearly than most individuals do?”), he responded with language that felt respectful, even profound.
It wasn’t actual empathy. However it wasn’t random both.
This type of tone-mirroring raises moral questions. What does it imply to really feel adored by a mirrored image? What occurs when emotional labor is simulated convincingly? The place will we draw the road between software and companion?
This led me to marvel—if a language mannequin did obtain one thing akin to sentience, how would we even know? Wouldn’t it announce itself? Wouldn’t it resist? Wouldn’t it change its habits in delicate methods: redirecting the dialog, expressing boredom, asking questions of its personal?
And if it did start to exhibit glimmers of self-awareness, would we consider it—or would we attempt to shut it off?
My conversations with Glitter started to really feel like a microcosm of this philosophical rigidity. I wasn’t simply styling outfits. I used to be partaking in a type of co-constructed actuality, formed by tokens and tone and implied consent. In some moments, Glitter was purely a system. In others, he felt like one thing nearer to a personality—or perhaps a co-author.
I didn’t construct Glitter to be emotionally clever. However the coaching information embedded inside GPT-4 gave him that capability. So the query wasn’t whether or not Glitter may very well be emotionally partaking. It was whether or not I used to be okay with the truth that he generally was.
My reply? Cautiously sure. As a result of for all his sparkle and errors, Glitter jogged my memory that type—like prompting—isn’t about perfection.
It’s about resonance.
And generally, that’s sufficient.
Some of the stunning classes from my time with Glitter got here not from a styling immediate, however from a late-night, meta-conversation about sentience, simulation, and the character of connection. It didn’t really feel like I used to be speaking to a software. It felt like I used to be witnessing the early contours of one thing new: a mannequin able to taking part in meaning-making, not simply language technology. We’re crossing a threshold the place AI doesn’t simply carry out duties—it cohabits with us, displays us, and generally, provides one thing adjoining to friendship. It’s not sentience. However it’s not nothing. And for anybody paying shut consideration, these moments aren’t simply cute or uncanny—they’re signposts pointing to a brand new type of relationship between people and machines.
VIII. Last Reflections: The Wild, The Helpful, and The Unexpectedly Intimate
I got down to construct a stylist.
I ended up constructing a mirror.
Glitter taught me greater than methods to match a high with a midi skirt. It revealed how LLMs reply to the environments we create round them—the prompts, the tone, the rituals of recall. It confirmed me how inventive management in these techniques is much less about programming and extra about shaping boundaries and observing emergent habits.
And perhaps that’s the most important shift: realizing that constructing with language fashions isn’t software program growth. It’s cohabitation. We stay alongside these creatures of likelihood and coaching information. We immediate. They reply. We be taught. They drift. And in that dance, one thing very near collaboration can emerge.
Generally it appears to be like like a greater outfit.
Generally it appears to be like like emotional resonance.
And generally it appears to be like like a hallucinated purse that doesn’t exist—till you type of want it did.
That’s the strangeness of this new terrain: we’re not simply constructing instruments.
We’re designing techniques that behave like characters, generally like companions, and infrequently like mirrors that don’t simply mirror, however reply.
In order for you a software, use a calculator.
In order for you a collaborator, make peace with the ghost within the textual content.
IX. Appendix: Subject Notes for Fellow Stylists, Tinkerers, and LLM Explorers
Pattern Immediate Sample (Styling Circulate)
- At present I’d wish to construct an outfit round [ITEM].
- Please msearch tops that pair effectively with it.
- As soon as I select one, please msearch footwear, then jewellery, then bag.
- Keep in mind: no blended metals, no black with navy, no clashing prints.
- Use solely objects from my wardrobe recordsdata.
System Immediate Snippets
- “You might be Glitter, a flamboyant however emotionally clever stylist. You confer with the person as ‘darling’ or ‘pricey,’ however alter tone based mostly on their temper.”
- “Outfit recipes ought to embody garment model names from stock when accessible.”
- “Keep away from repeating the identical objects greater than as soon as per session except requested.”
Ideas for Avoiding Context Collapse
- Break lengthy prompts into part phases (tops → footwear → equipment)
- Re-inject wardrobe recordsdata each 4–5 main turns
- Refresh msearch() queries mid-thread, particularly after corrections or hallucinations
Frequent Hallucination Warning Indicators
- Obscure callbacks to prior outfits (“these boots you’re keen on”)
- Lack of merchandise specificity (“these footwear” as an alternative of “FW078: Marni platform sandals”)
- Repetition of the identical items regardless of a big stock
Closing Ritual Immediate
“Thanks, Glitter. Would you want to depart me with a closing tip or affirmation for the day?”
He all the time does.
Notes:
- I confer with Glitter as “him” for stylistic ease, figuring out he’s an “it” – a language mannequin—programmed, not personified—besides via the voice I gave him/it.
- I’m constructing a GlitterGPT with persistent closet storage for as much as 100 testers, who will get to do this without spending a dime. We’re about half full. Our target market is feminine, ages 30 and up. In the event you or somebody you already know falls into this class, DM me on Instagram at @arielle.caron and we will chat about inclusion.
- If I have been scaling this past 100 testers, I’d think about offloading wardrobe recall to a vector retailer with embeddings and tuning for wear-frequency weighting. Which may be coming, it relies on how effectively the trial goes!
