Vertical pack · Beauty
The beauty pack is a file you can count.
Hair care and makeup, and with them skin care, beard and shaving care, and fragrance. It is the one pack that exists today, and it is a data file: the questions it asks, the sentences it refuses and the names it falls back to are all entries you can be shown.
- Products
- Variants & options
- Prices
- Live stock
- Collections
- Order status
- Returns
- Store policies
- Answers you write
- A code you choose
A pack changes what Refik asks; it leaves the ten sources an answer is built from exactly where they were. Every store reads the same ten, shampoo or otherwise — what Refik reads.
What changes
What changes the moment this pack loads.
Two things, and both of them are narrowings. The pack gives the assistant a shorter list of things it is allowed to say, and a better idea of what to ask first.
Three questions, and eight answers that name a need
Hair type, main concern, heat use — four, seven and three answers. Eight of the fourteen carry the name of a need, and that name has to be one of the twenty-eight the pack itself defines: an undefined need is refused when the file loads, so every live option points at a tag your catalogue can carry. The other six carry a label only, on purpose, because “fine hair” describes hair while your shelves hold products. What the shopper finally sends is a three-slot template from the file, filled with their own words.
Twenty-four sentences it refuses, three of them in one market
The rules are read against the draft answer before the shopper sees it. Twenty-one of them apply wherever the store sells. Three carry a market, and all three carry the same one: a “free-from” line the cosmetics regulator there forbids is the name of a real collection somewhere else. Attaching the market to the rule is what lets one file be right in both places.
Counted from the file
Every number here was counted from the file.
The pack is a data file with eighteen top-level fields. Four of those fields belong to this pack alone and are absent from the neutral one: the advisor description for the profile tool, the examples of what usually goes wrong, the storefront block, and the short phrases the assistant is allowed to use about what a product does.
46
fallback names for the shelves a beauty store keeps, from shampoo to brow gel. Where your own catalogue already names a tag, your name wins.
28
needs a product can answer — dryness, frizz, heat protection, thinning hair — each one a tag your catalogue may already carry.
13
patterns that mark a question as one for a person. Refik says so on the spot and offers to pass the conversation to you, with everything the shopper wrote.
50 alternative spellings
Nine of them carry Turkish letters; the rest are Spanish. A shopper who types the Spanish word for a mask, a conditioner or a wax is understood while your catalogue stays in its own language, and the three languages here are the three a claim rule can be written in.
39 ingredient terms
Argan, keratin, biotin, collagen, quinoa and on down the list. They are read while the product knowledge is built, so an answer can say which of your products carries which — read out of your own ingredient text.
13 replenishment intervals, 12 to 50 days
Per role, per hundred millilitres, which is why a shampoo and a beard oil come up on different schedules. Every interval is arithmetic from size, and the window around it — no reminder before twenty-one days, none after two hundred and forty — lives in the code and is the same for every pack.
46 storefront entries, each in three languages
The greeting, the example questions, the survey label, the entry cards, the bubble prompts. They are written to app-owned metadata when you confirm the pack. Where a pack has an empty storefront block the generic wording stays, and the same write clears the text it replaces.
Three switches, all three open
Shade matching, a remembered shopper profile, and noticing a product that is running out. Two of the three are tools the model is only told about when the pack opens them; the third is a list of reminder candidates that comes back empty when it is closed.
5 profile fields
Hair type, colour-treated, porosity, heat use, and what bothers you most: two free text, one boolean, two from a short list. A sixth field is refused the moment the assistant tries to write one, and the refusal comes back with the list of the five it has.
Nine counts, read out of one file and the code that reads it. The survey’s own intro and its sentence template are written in two of the three languages and fall back to English for the third; the storefront block is written out in all three. English is the one language a pack always fills, because every blank falls back to it.
What this page is a picture of
scripts/sektor-maket-yaz.py.Counted
Four more counts out of the same file.
These four describe the parts of the pack a shopper meets last: how a question is routed, what the assistant may say about a product, which questions come to you, and where the threshold came from.
-
Four intents, matched before the model runs
A colour chart, shipping and tracking, returns and exchanges, and how to use a product. Each is a pattern in the file, in English and Turkish, and a question that matches one is routed to the tool for it rather than composed as prose.
-
Three phrases about what a product does
Hydrates, conditions stressed strands, helps reduce breakage — in English and Turkish. That short list is the whole vocabulary the pack allows for a product benefit, which is why an answer sounds specific and stays inside the twenty-four claim rules.
-
Thirteen handover patterns of its own
Above the base nine: a named skin condition, a recent procedure such as a peel or filler, an acne treatment or retinoid, contact lenses, an eye infection, pregnancy or breastfeeding, and an allergy — written across English, Turkish and Spanish. Those thirteen come to you with the conversation attached.
-
Three catalogues behind the threshold
The threshold for suggesting this pack was measured, not chosen by taste. In two beauty catalogues, 139 of 142 products and 138 of 141 carried at least one of the pack’s words; in an unrelated catalogue of furniture and electronics the count was zero out of ten. Both beauty catalogues are our own, so this is a calibration of the threshold.
Unchanged
What stays yours, under every pack.
These three are the same on every vertical page on this site, and that is the point: they describe the part of Refik that your store decides and a pack only reads.
You choose it, and the server measures again
Refik counts how many of your products carry any of a pack’s words and shows you the count, the ranked score, up to eight of the words it matched and up to three example products for each. A pack is only suggested when the share clears a threshold, at least three different words matched and at least five products matched. You confirm it, and the server runs the measurement again on its own before it writes anything. How a pack gets chosen.
Your catalogue supplies the products
A pack supplies names; your catalogue supplies the products. Every option Refik offers has passed two gates in your own store — sellable, and in stock — so the products a shopper sees are the ones you carry today. A pack tag that matches zero products in your catalogue is counted and shown in the console, with the number it matched.
The base set stays merged in
Sixteen claim rules and nine handover patterns are merged in when any pack loads, in English, Turkish and Spanish, and they sit behind the pack’s own rules so the pack’s rule is still the first one reported. A base rule that would report the same words twice stays quiet. An empty set makes the app refuse the pack outright. The base set.
A store outside those five gets the neutral pack — what it still does, and what it narrows. Both pages name an industry only where a file for it exists.