selves.me
A private journal that quietly builds a living identity portrait from your own words. Only when you're ready does any of it go public.
Every tool built to help people express themselves online starts in the wrong place. Substack opens to a blank page and asks you to write a newsletter before you know what you want to say. LinkedIn asks you to describe yourself before you know who you are. Linktree asks you to curate a presence before you have one. The blank public page is the enemy of honest self-articulation. Most people carry the question who am I, really? quietly for years without ever getting closer to an answer.
Invert the whole pattern. Start with a private journal: thirty seconds of voice on the way home, two lines before bed, a photo of something that mattered today. No audience, no performance, no pressure to sound interesting. As entries accumulate, an AI reads everything the user has written and quietly builds a living model of who they actually are: their values, their patterns, the story they're currently in, the gap between what they believe and how they live. Only then, and only if the user wants, does any of it go public, as a living identity portrait written in their own voice.
- Multi-modal capture: voice (transcribed live via native speech recognition), text, and photos; the thought is always saved before AI processing begins
- Capture reflection: Claude Haiku returns a 1–3 sentence observation on each capture immediately; observational only, never therapeutic, never a question
- Worldview synthesis: Claude Sonnet reads all captures in chronological order and generates a structured narrative portrait using NT Wright's four-question framework; gated behind a free-tier limit
- Praxis privacy: the most sensitive output (the gap between stated values and actual behavior) is enforced private at three independent layers: excluded from the embedding used for discovery, stripped from the API response before it leaves the server, and fetched only via explicit user opt-in
- Public identity page at selves.me/slug: published only by the user, only when ready; RLS rejects the write at the database level if the user isn't on Pro
- Open Graph injection without SSR: Edge Middleware intercepts crawler requests, fetches live page data, and injects meta tags before serving; the browser still gets the SPA
- Dual-payment architecture: Stripe on web, StoreKit 2 on iOS through a first-party Capacitor plugin with no third-party SDK; every Apple purchase is verified server-side against the App Store Server API, and both paths resolve to one tier the client reads as a single value
- Founding member tier: auto-granted at signup for early users via a database trigger, no application logic required
Live on the App Store and at selves.me. Captures accumulate, portraits synthesize, founding members are in.
React 19 · Vite · React Router · React Query · Vercel
CSS custom properties + CSS Modules (no utility framework)
Supabase: Postgres, RLS, Storage, Edge Functions (Deno)
Claude Haiku: immediate capture reflection; latency-sensitive
Claude Sonnet: structured JSON worldview portrait; NT Wright framework
OpenAI embeddings: semantic vectors for portrait similarity search
Web Speech API (browser) · native speech plugin (Capacitor/iOS)
Stripe (web) · StoreKit 2 via first-party plugin (iOS), both write to one subscriptions table
Capacitor: React SPA in a native shell; handles deep links and IAP
PostHog
Three decisions that don't show up in a feature list. Praxis privacy: the most sensitive AI output is enforced private at every layer it travels: excluded from semantic embeddings, stripped from the API response before it leaves the server, and only fetched when the user explicitly requests it. Tier enforcement at the database: the RLS policy on public pages rejects unpaid publish attempts at the Postgres level, not the UI. No client-side bypass, stale cache, or missing paywall check can override it. Multi-payment coherence: Stripe and StoreKit are two separate billing systems writing to the same subscriptions table. Neither app can grant Pro itself: a Stripe webhook and a server-side Apple check are the only paths in. A single hook on the client resolves all payment sources, including the founding member tier, into one consistent Pro state. Scattered entitlement checks across components is how you accidentally give things away.
Selves is the clearest case for taste is the moat. The blank-page inversion was an opinion, not a feature. But the more transferable argument is architectural: every hard constraint in Selves (RLS on every table, secrets only in edge functions, tier enforcement at the database, praxis privacy across three layers) was made once and never revisited. That left room to move fast on prompts, copy, and UI. Treating architecture decisions as irreversible and surface decisions as cheap is the right order of operations for any AI product. That's what Barael sells.
Live on the App Store and at selves.me. Captures accumulate, portraits synthesize, founding members are in.