A product design case study for an AI palmist rooted in the shastras, built to name the line it read.
Palm reading online sits between two bad experiences, and both fail the person holding the phone.
The design brief wrote itself from three honest problems. Readings feel generic, because most tools read from a tiny set of features and fill the rest with cold-reading. The money is made through dark patterns the FTC has flagged, the kind that quietly empty a wallet. And a palm photo feels more biometric than a birthday, yet nobody in the category treats it that way.
Most palm apps guess. Ours can name the exact line.The product thesis, in one sentence
I ran the product on two research tracks at once: the market it competes in, and the tradition it claims to stand on.
Market teardownI pulled apart the incumbents. Co-Star and The Pattern for their minimal, brandable craft. AstroTalk, where a large share of revenue is marriage anxiety. Nebula and the palm-scanner cluster for the exact dark patterns to avoid, expiring credits, timer-delayed close buttons, gate-to-see-results. The teardown produced a list of moves the category runs, and the specific gaps an honest product could take.
The canon studyThen the harder question. Why does every AI reading sound identical? The answer was structural, not stylistic: a reading built from hand shape plus the three major lines is drawing from a tiny feature space. Give the model more real vocabulary to see, and the sameness dissolves.
So I commissioned a deep study across six living traditions of the hand: Samudrika Shastra, Western chiromancy, Chinese five-element hands, the Perso-Arabic ilm al-kaff, Japanese teso, and the Tamil Nadi thumbprint taxonomy, plus modern dermatoglyphics. It became a written report, The Shastras of the Hand, and it carried its own honesty flags: the popular Earth, Air, Fire, Water hand types are a 1965 invention and must never be sold as ancient, and pseudoscience like DMIT and finger-ratio claims were marked radioactive and kept out.
A reading that cannot point to the feature it springs from is a fabrication, and you do not fabricate.The rule that governs every line of output
The tradition is not decoration here. It is the data model, the vocabulary, and the trust.
Every reading is grounded in Hindu Samudrika Shastra, the hast-rekha, enriched by Western chiromancy after Cheiro and Benham. The sources are cited, not implied.
The palmistry corpus grew from a first draft to over one hundred thousand characters across fact-checked domains: the major and minor lines, mounts, hand shape, fingers, thumb and nails, fingerprints, and the Samudrika special signs. It loads into the model as one cached block of ground truth.
In Samudrika the primary hand is the right for a man, the left for a woman. But telling left from right on a palm held flat to a camera is genuinely unreliable, so the product does not guess it. It asks, then treats your answer as truth. A tradition kept, without a lie of confidence.
Decision: ask rather than detect, because a wrong guess here breaks trust in everything after it.Seven themes carry the reading. Underneath them sits a fixed vocabulary the model must work in: thirteen lines, eight mounts, and the classical special signs, the star, the trident, the island, the mystic cross. Naming the parts is what lets the product cite its evidence later.
The category looks like a carnival. I designed the opposite: a quiet, old-book stillness that signals care.
The world is a night sky in plum, with a single warm parchment card holding the reading, edged in gold. Display type is Marcellus, the reading is set in serif faces, Cormorant Garamond and Spectral, with wide-tracked gold labels. Motion is restrained: stars twinkle, lines draw themselves on with a slow stroke, and everything respects reduced-motion. Gold is spent sparingly, on the one thing that matters in each view.
The signature artifactEvery reading renders a personalised diagram of the hand, the lines and the eight mounts in their real anatomical places, drawn to what the AI actually saw in your photo. The styling is the honesty:
A line the model could not see is drawn as a faint parchment dash, the map learned, not claimed. Show a left hand and the whole drawing mirrors, while the labels counter-mirror so every glyph stays upright. The reading also ships as a printable PDF that bundles fonts for eleven Indic scripts and Arabic, so the report reads correctly in the user's own language.
For Him, For Her, or For Two, each carrying the Samudrika hand note so the tradition is taught in one line, not hidden.
The second biggest conversion killer is a bad photo. So the do and do-not guide, flat hand, fingers apart, even light, appears before the camera, not as an error after a sixty second wait. A live alignment overlay helps, with a graceful fall back to upload.
A client-side check flags a photo that is too small, too dark, or too blurry, before spending an expensive AI call. It is deliberately lenient, because a false reject is worse than a borderline photo, and it says so: this is a quick technical check, the full reading still checks the rest.
The free reading is a complete, satisfying mini-reading, three full themes, the hand's element, and a taste of the luckiest mark. What it holds back is deliberate, covered next.
Anyone can call a vision model. The product is the set of decisions that make a person believe, and correctly doubt, what it says.
Every theme carries the exact feature it was read from, shown as "read from" the lines and mounts behind it. A collapsible panel, "the signs behind it", opens the full reasoning. The output schema forces the model to declare, per line, whether it saw it clearly, faintly, or not at all, so it cannot quietly invent what it could not see.
Why: trust in AI is not built by confidence, it is built by showing the work and admitting the gaps.A reading takes about a minute. Rather than a spinner, the screen narrates the craft: steadying your hand in the light, tracing the heart, head and life lines, weighing the mounts, cross-referencing the two traditions, composing your reading. The subtitle: the lines are speaking.
Why: a slow AI response, framed as care, reads as depth rather than lag.Point it at a diagram or a cupped hand and it declines instead of hallucinating a reading, with warmth, not reproach. When something breaks, the error is in character: the lines stayed quiet.
A whole layer of the prompt is titled read this hand, not a hand, with specific banned cold-reading phrases. A standing eval harness scores every reading for genericness and distinctiveness with an independent judge, and the score is tracked release over release. Sameness is treated as a measurable bug, not a vibe.
Why: the one thing the whole category gets wrong is the one thing we measure.Readings are framed as tendencies and leanings, never fixed fate. Health is a strict boundary, coded and eval-enforced: never medical, never a prediction of lifespan or death, the life line never read as length of life. The category's easiest revenue, anxiety, is refused on purpose.
"Ask the reader" is a follow-up chat, grounded in the frozen reading so it never contradicts itself or conjures a line that was not there. "Speak to me" narrates the reading aloud with word-synced highlighting, and any reading translates into the user's language in a native palmist's voice, not translated English.
The most personal thing a user hands over is a photo of their body. The design makes that safe, and says so at the exact moment of hesitation.
The photo is deleted the instant the reading is done, never stored, never sold, never used to train. No face, no fingerprint, no biometric is created or kept. There are no ad networks and no tracking. This promise is repeated on the upload screen, in the consent banner, on the report, and on the share card, because the upload screen is where doubt peaks. Every reading carries an honest disclaimer that it is symbolic, not medical or financial advice, and the paywall is backed by a plain money-back guarantee.
We read your palm. We do not sell it.A privacy stance used as a marketing wedge
The hardest constraint was making money the way the category does not.
The free reading is genuinely complete, three full themes chosen on purpose to be the lower-intent ones. The higher-intent themes, marriage, wealth, and the cautions, are what the detailed report unlocks, along with every line and mount, the illustrated PDF, audio, and the emailed report. A running counter shows real progress, you have revealed three of seven chapters, and nothing in the free teaser gives away what sits behind the wall.
No hidden auto-renew, no gate-to-see-results, no expiring credits, no timer-delayed close button. The things the category leans on are the exact things banned in the brief.
The intro price holds until a real milestone of readings, then steps up on its own. Scarcity that is actually true: intro price, ends at five thousand readings.
Beyond the single reading: Saath, a two-palm compatibility report, a twelve-month Year Ahead forecast, a monthly guidance subscription grounded only on the stored reading, and a family credit pack. Each priced in clean round numbers, resolved to the user's currency.
One person, one coherent product, across every surface where a decision was made.
I owned the whole arc. The market and canon research. The information architecture and the seven-theme model. The visual system, tokens, type, and the palm-map illustration. The microcopy, every line the reader sees. The AI experience, how it earns trust, handles latency, refuses gracefully, and stays honest. The paywall logic and the ethics rails. And uniquely for an AI product, the prompt itself was a design surface, the voice, the banned phrases, the grounding rules, all authored as one piece with the interface. Copy, code, config, pricing, ethics, and research, held together as a single product.