Prestolex — the consumer web application

Prestolex (github.com/udopnink/nai-prestolex) is the third web application in this portfolio and the first aimed at consumers rather than at Udo's own consulting practice. It modernizes the mechanic of a classic parlour word race — a category and a letter are revealed, everyone races to type a fitting word — and replaces the argument the paper game always ended in with an AI referee that rules every answer instantly, with a reason. It took over as the primary commercial project on 2026-08-11, inheriting the previous candidate's build envelope, and shipped its first live milestone within days.

Like nai-analysis and nai-onto, it is built with GSD's phase-based discuss→plan→execute→verify loop rather than this workspace's orchestrator/subagent/tool pattern — a persistent product with a database, a deploy pipeline, and now a billing system, not a one-shot dispatch that produces a document. Unlike the other two, an LLM sits inside the product: the referee and the deck generator are Claude calls at runtime, which is why this app has a model policy and cost guardrails the other two never needed.

The game

The AI referee

Cost guardrails, named before anything can go viral

Every limit is env-configurable, but the point is that each exists and has a name: a per-account monthly judgment allowance (300 free, 8 000 plus, credit packs beyond that), per-room daily caps enforced atomically with the allowance, three deck generations per device per day on the free tier, and a global spend alarm — soft at €10/day, hard stop at €25/day. Exhaustion and the hard stop don't error: they flip judging to majority-vote mode, the same graceful floor the referee falls to when it's unavailable.

Accounts, caps, and the paywall

How it's built

%%{init: {'theme':'neutral', 'flowchart':{'htmlLabels':false,'nodeSpacing':45,'rankSpacing':55}}}%%
flowchart TD
    H["The console - TV, tablet or laptop<br/>host signed in - the single writer"] --> API
    P["Players' phones - anonymous<br/>QR join - submit answers only"] --> API

    subgraph L1 ["Next.js server routes - api/rooms/*"]
        API["applyActions over the pure reducer<br/>version-guarded on rooms.version"]
        REF["Referee module - never throws<br/>memo hit? no API call"]
    end

    API --> REF
    REF --> CL["Claude API<br/>claude-haiku-4-5, structured output"]

    subgraph L2 ["Supabase - service role only for writes"]
        DB[("rooms · answers (append-only)<br/>decks · subscriptions · allowances")]
        MEMO[("judgment memo<br/>each distinct answer judged once")]
        RT["Realtime channel per room<br/>broadcast + presence"]
    end

    API --> DB
    REF --> MEMO
    DB --> RT
    RT --> H
    RT --> P

    classDef orch fill:#dcefe9,stroke:#0f6f63,stroke-width:2px
    classDef tool fill:#f4f1ea,stroke:#8a7a5a
    classDef human fill:#fdf3e0,stroke:#8a5a00,stroke-width:2px
    class H,P human
    class API,REF orch
    class DB,MEMO,RT,CL tool

The seam to this workspace

Like nai-onto, and unlike nai-analysis, Prestolex has no seam to this workspace at all — no slash command reads it, writes to it, or knows it exists. What it shares with the agent systems is only the discipline: pinned decisions that don't get relitigated, invariants with a named single home in the code, degraded modes designed as features, and a verify step before anything is called done.

Direction