AppHub

A read-only, skill-first knowledge base for mobile-app & SaaS growth.

AppHub holds ~268 bite-size, evidence-backed entries — each a single claim with verbatim numbers, a confidence level, and its source — grouped by topic (onboarding, paywalls, pricing, retention, ads, ASO, attribution, …). It's exposed as a JSON API and a Claude Code skill so an AI assistant pulls from it automatically when you hit a growth problem, grounding its advice in real data instead of generic tips.

Live: apphub.marijn.fr

The skill is the product

The primary consumer isn't a human browsing — it's your AI assistant:

You ask your assistant a growth question
        │
        ▼
  GET /api/catalog      → ingest the whole catalog, reason over the claims
        │
        ▼
  GET /api/search?q=…   → synonym-expanded search for anything else
        │
        ▼
  GET /api/wiki/<cat>/<slug>   → read the entries that matter
        │
        ▼
  Read your app's real code (onboarding, paywall, pricing, …)
        │
        ▼
  Prioritized, cited fixes — each backed by a sourced entry

The web UI (/library) is a clean way for humans to browse the same content; install instructions live at /connect-your-ai-agent.

Content model

Two types, both markdown with YAML frontmatter (full spec: docs/CONTENT-FORMAT.md):

• Entry — one atomic insight. claim, numbers, topics, confidence (high / medium / debated), source. A few hundred words; Apply when / Caveat.
• Guide — the rare long-form keeper (e.g. the RevenueCat SOSA 2026 benchmarks).

Contradictions between sources are never silent — they're tagged confidence: debated with both sides in the caveat. Sources are cited explicitly (App Society calls, RevenueCat, Marc Lou, named guides).

Where the content comes from

• App Society — a weekly French indie-app mastermind (12 calls distilled). Fed automatically: the app-society-sync skill turns each new call into entries and pushes them here.
• Marc Lou viral-product principles, RevenueCat State of Subscription Apps 2026, and a few sourced practitioner playbooks.

API

All content is JSON. Start at /api/catalog.

GET /api/catalog                 # compact catalog of EVERY entry (read this first); ?topic=<name> to filter
GET /api/search?q=onboarding     # synonym-expanded + stemmed full-text search, with snippets
GET /api/wiki/<category>/<slug>  # read one entry (markdown + metadata)
GET /api/wiki/nav                # topic tree
GET /api/og/<category>/<slug>    # per-entry Open Graph image
GET /api/help                    # API overview + agent system prompt

/api/catalog is served from a precomputed meta/catalog.json (a single read — reading every entry per request times out on serverless at this size).

Read-only

AppHub is read-only in production. The legacy submission / Q&A-generation endpoints (/api/resources, /api/queue, /api/ask, article creation) return 403 unless ENABLE_WRITES=true. Content is curated offline and deployed via the scripts below.

Architecture

Next.js 16 (App Router) · Cloudflare R2 (storage) · Vercel (hosting)

src/
  app/
    library/            # browse all entries (one /api/catalog call) + entry/topic pages
    connect-your-ai-agent/  # skill install + API docs
    api/
      catalog/          # cached compact catalog
      search/           # MiniSearch + synonyms + stemming
      wiki/[...slug]/    # read an entry
      wiki/nav/          # topic tree
      og/[...slug]/      # per-entry OG image
      help/ index/       # API docs / full index
  lib/
    r2.ts               # storage (R2, or local ./data fallback)
    search.ts           # index + catalog builder
    nav.ts              # topic tree builder
    synonyms.ts         # query expansion
    topics.ts           # topic labels
    markdown.ts         # frontmatter parsing

Storage layout (R2 / ./data):
  wiki/<topic>/<id>.md       entries & guides
  wiki/<topic>/_meta.json    topic title/order
  meta/{catalog,search-index,nav-tree}.json   precomputed, rebuilt on deploy

The web route is /library but the storage prefix and data API stay wiki/ / /api/wiki (/wiki + /developer 308-redirect to the new routes).

Scripts

npx tsx scripts/push-entries.ts    # incremental: upsert data/_incoming/** to R2, rebuild from R2 (safe, never deletes)
npx tsx scripts/deploy-to-r2.ts    # full rebuild from local data/wiki/** (refuses if the mirror looks stale; --force to override)
npx tsx scripts/reindex-local.ts   # rebuild index/catalog/nav locally (./data)
npx tsx scripts/ls-bucket.ts       # inspect what's in the bucket (--keys, or pass a key to dump it)

Local development

npm install
R2_ENDPOINT= NEXT_PUBLIC_AI_ENABLED=false npm run dev   # uses local ./data, no R2/AI

Set R2_* in .env.local to target the real bucket. The site is fully functional without AI — browse, search, and the API all work statically.

License

MIT