JMTarot

Jodith Miftah Tarot — a mobile-first tarot website, installable to the iPhone home screen. Three reader personas, three services, the 22 Major Arcana. Every reading is written by a model at request time, in Indonesian or English, in the voice of whichever reader you picked.

www.jmtarot.site · v0.8.0 · Next.js 16 on Vercel sin1

pick a reader ──► pick a service ──► draw 3 of 22 ──► POST /api/reading
                                          │                   │
                          your question ──┤        moderation gate (glm-4.5-flash, 1.5s)
                                          │                   │
                    the cards that keep ──┤          glm-4.6, streamed, four paragraphs
                       coming back to you │                   │
                                          ▼                   ▼
                                    /history · /s/<slug> · the group chat

---

Demo

Every capture below is the running app at 390 x 844 — an iPhone 12/13/14/15 width — at 2x, driven over CDP against npm run dev, a real Postgres and real z.ai calls. The readings, the chat turns and the translations were all generated at the moment of capture; the reader portraits and the card art are the only pre-existing images.

---

What it does

• A reading, generated for you, now. Pick Thessaly, Margaret or Adrian; pick a daily card, a three-card spread or a yes/no; type a question or leave it blank; draw from a fan of 22. The prose streams in that reader's voice, in your language, and is stored.
• The answer is derived in code, not chosen by the model. effectiveYesNo() decides yes / no / unclear from the card and its orientation — reversal flips it — and the prompt is handed the word. A question offering options ("ayam atau ikan?") gets the same treatment from the other end: the model picks, and code validates the pick is a slice of the question you typed.
• It remembers, and says so without doing arithmetic out loud. A written verdict on the cards that keep arriving, a reading that refers back to your last one, and a per-day summary in your reader's voice. The model is never given a count — it gets two card names, a Shadow Arcana and a dominance word, so it cannot recite a tally it was never told.
• Bilingual, interface and readings. id and en. Switching does not regenerate anything: it translates the stored prose, checks the card names came through mechanically, and caches the result. Your question is never translated, on any surface.
• A group chat that keeps going when nobody is looking. All three readers and you in one room. A director writes a beat sheet, the voices execute it one turn at a time, and the readers can open a conversation unprompted. Silence is a valid outcome — there is no error bubble, because a friend who has nothing to add says nothing.
• Onboarding once, then never asked again. Nine screens, six free-text answers, AES-256-GCM encrypted at rest, distilled into a "Lotus" summary that grounds every later reading. A skipped answer stays skipped, and the readers never ask about the thing you declined to answer.
• Sharing that a stranger can open. /s/<12 chars> — no account, no indexing, revocable, and it carries the language it was shared in. Re-sharing rotates the slug within that language; revoking kills every language at once.
• A public half, for people with no account. A signed-out landing page, a gallery of all 22, a lore page per card, 44 wallpaper derivatives, and a blog with a markdown editor behind it. 54 indexable URLs across two language trees.
• A moderation gate that refuses harm without refusing tarot. Grief, illness, money trouble, divorce and a frightening partner are the subject matter — they are not refused. Eight categories where the answer itself would be the harm are, and the refusal is never delivered in a reader's voice.
• An admin panel that measures and never restrains. Thirteen panels, token cost per operation, a model-written finding under each chart, and per-user detail with counts and no text.

Not built, deliberately: birth card, the once-per-day card lock, an About page. Card names stay English in both locales, and so do the readers'.

The three readers

	Voice	Specialities
Thessaly · The Grounded Guide	intuition plus logic, plain and lit	career decisions, direction, untangling a problem
Margaret · The Old Soul	long sentences, 1.3x everyone else's word budget	self-discovery, inner reflection, family
Adrian · The Modern Mystic	short, contractions, will make a joke	love, self-worth, short-term decisions

Each carries a fixed gender in readers.json — Thessaly and Margaret female, Adrian male — because a gender does not vary by locale but the words do. Each persona block in the prompt carries one worked example paragraph, and that paragraph does more work than the description: if the three ever stop being distinguishable with the names covered, the fix is there and not in the code.

npm run smoke -- --all prints eighteen readings and ends with a blind read — three per locale, names covered, key after forty blank lines. Guess. Three of three or the personas need work.

This is the beginning of the README. Read the rest at the original. Read the full README ↗