How the advisor actually works.
Not a buzzword-laden pitch. An honest, technical explanation of what happens when you take the quiz — what the AI reads, how it decides, when it refuses, and why the whole thing is free.
What happens when you take the quiz
You answer eight questions about your age, skin type, primary concerns, current routine, sensitivities, life context (pregnancy, perimenopause, prescription skincare, recent weight changes), monthly budget, and anything else you want us to know. Those answers are sent to Claude — the AI model built by Anthropic — along with a system prompt we wrote that constrains what Claude is allowed to recommend and what safety rules it must follow.
Claude reads your inputs, considers our catalog of pre-approved products, and generates a routine — typically a small set covering cleanser, treatment, moisturizer, and SPF. For each one, it writes a plain-English reason explaining why that product suits your specific profile, lists the active ingredients, and shows you what we earn (or don't earn) if you buy it.
Then it does one more thing: it lists 2–3 products it considered but rejected, with brief reasons why. That last bit is intentional. You deserve to see not just what the AI picked, but what it chose not to.
Then it builds your routine. Not just a list of products — an actual day-by-day plan: a morning sequence and an evening sequence, with each step in order, specific application instructions, and timing. If your routine includes a retinoid or other active that needs ramping up, the AI writes a first-month schedule explaining how to introduce it gradually and what to expect. It also adds lifestyle notes where they matter — SPF reminders, sun exposure cautions during retinoid use, anything that affects whether the routine actually works. The whole thing lives on your dashboard, ready to follow.
What a report actually looks like
Two weeks after your baseline photo, you can generate your first AI Progress Report. This is a real rendering of what appears on your dashboard — built using the same code, showing sample data from a hypothetical 6-week check-in with a combination-skin routine. Take a look:
Your Week 6 report
Six weeks in, the niacinamide is clearly doing its job — redness is down, tone is more even. Texture hasn't shifted yet, but that's expected at this point in a retinol routine. You're on-curve.
Both photos taken in natural light around the same time of day. Angles are comparable. Minor lighting warmth difference in the latest photo may slightly exaggerate the luminosity read.
Niacinamide tone-evenness effects typically emerge at 4–8 weeks — your results are tracking with that timeline. Retinol texture effects typically emerge at 8–12 weeks, which is why no texture shift is visible yet. Give it another 2–4 weeks before reassessing.
The routine is producing the changes it should at this point. The niacinamide is working, the retinol is still in its ramp-up window, and there are no concerning observations. Stay consistent — take your next photo in 2 weeks.
The five principles, in practice
The five commitments on the homepage aren't marketing copy — they're rules encoded into how the AI operates. Here's what they mean concretely.
I. We explain our reasoning. Every recommendation ships with ingredient-level logic: not "this is great for dry skin," but "hyaluronic acid and ceramides together restore the lipid matrix your barrier is missing, which is why your skin feels tight by noon." If our reasoning seems wrong, you can challenge it — the system prompt instructs Claude to defend its picks with mechanism-of-action detail.
II. The AI watches it work. Every two weeks, if you upload a new progress photo, Claude compares it to your baseline and gives you an observational read: what's visibly changed, what hasn't, whether lighting/angle differences are confounding the comparison, and a verdict of "stay the course" or "consider adjusting." It will tell you when it can't tell — when the photos are too different to compare fairly. That honesty is the point.
III. We publish our commissions. Next to every product, you'll see exactly what percentage we earn if you buy through our link. On our non-affiliate picks (Cetaphil, Vanicream, Bioderma, Avène, Aquaphor, and most of CeraVe), it's zero. You can factor that into how you weigh our advice — and you should.
IV. We pick products we don't earn on. Roughly one in every five recommendations pays us nothing. Our system prompt requires it. These picks exist because sometimes the right answer is the $14 drugstore moisturizer, and pretending otherwise — just to protect our commission — would make us exactly the kind of site we were built to replace.
V. We refuse when we should. If your answers suggest an irritated barrier, an active reaction, or a situation where no catalog product is safe, the AI is instructed to say so explicitly and recommend you wait, patch-test, or see a dermatologist. It will not invent a "soft" recommendation just to satisfy the request. We'd rather disappoint you with a refusal than harm you with a bad pick.
When we refuse
Specific cases where the advisor will decline or redirect rather than recommend:
Pregnancy. No retinoids, no high-concentration salicylic acid, no strong actives. The catalog narrows to pregnancy-safe options like azelaic acid, niacinamide, and barrier-repair moisturizers. If you ask for something stronger, the AI will explain why it's not safe during pregnancy and suggest waiting.
Active barrier damage. If your intake notes describe recent reactions, peeling, weeping, or extreme irritation, the AI will recommend only gentle cleanser and barrier repair — no actives, no exfoliants. Usually with a direct note: "See a dermatologist if this persists more than a week."
Medical conditions beyond cosmetics. Cystic acne, rosacea flares, seborrheic dermatitis, eczema, suspicious moles — all routed to medical care, not product recommendations.
Budgets that can't accommodate real efficacy. If your concerns require actives but your budget is under $30, we'll recommend what fits — but we'll also tell you honestly which parts of your routine can't be addressed at that budget tier, so you're not disappointed expecting retinol results from glycerin.
Is it really free?
Yes. You will never pay evenskin.ai anything. No subscription, no premium tier, no paywall after your third question, no "pro" upgrade. You can take the quiz monthly, upload progress photos for years, and read every recommendation without spending a dollar.
We make money only when you click through to buy a product we recommend, via an affiliate link. The retailer (Amazon, Sephora, whoever) pays us a small commission out of their own margin — not yours. The price you pay is the same price anyone else pays. And for our non-affiliate picks — at least one in every five recommendations — we earn nothing even if you buy.
This model has one honest limitation worth naming: if no one ever bought anything, we couldn't afford to keep running. So we do want our recommendations to be good enough that people buy them. But the one-in-five non-affiliate rule and the refusal rule are structural guarantees that we can't quietly slide toward "recommend whatever pays most" without breaking our own system.
Safety and limitations
The advisor is a cosmetic skincare tool, not a medical one. It cannot diagnose skin conditions, prescribe medications, or replace a dermatologist visit. Every recommendation includes a disclaimer about patch-testing and consulting a healthcare provider for anything that looks like a medical concern.
Progress photo comparisons are observational, not diagnostic. The AI describes what it sees in the pixels — it doesn't detect disease, measure anything clinically, or substitute for in-person evaluation. If a photo shows something that looks like a medical issue, Claude is instructed to flag it and recommend seeing a professional.
Your photos stay encrypted and tied only to your account. We don't use them to train AI models. We don't share them with brands. We don't analyze them for demographics or sell aggregated insights. Any engagement data we look at is anonymized and aggregated — never photos, never identifying information.