The Unlisted

---

In one minute

The problem. Review apps reward the places that are already famous. The hole-in-the-wall spots locals actually love get buried under chains, ads, and star inflation.

What it is. A hand-picked guide to independent food and drink in the Bay Area, San Diego and Portland: 335 places, each with a one-line tagline, a story, a signature order, and a Hidden Gem Score. You collect stamps in a passport, let Surprise me deal you a place, or browse by neighborhood and time of day.

How it was made. AI did the heavy lifting and a person made every call. A pipeline screened 17,506 venues from open map data, filtered out chains, and had Claude (with web search) discover and research candidates. Every published place was reviewed by hand. Total AI research cost: about $70.

Built by Dan West, an AI strategy and operations leader with nearly 20 years in IT, by directing Claude Code as the engineering team.

Highlights

	For a visitor	Under the hood
Only independents	No chains, no ads, no paid placement.	OpenStreetMap data with a chain and landmark filter: 17,506 venues down to 12,630 independents.
An honest score	Five marks you can see, not a mystery star rating.	Uniqueness 25%, Atmosphere 20%, Story 20%, Signature item 15%, Under the radar 20%, computed from one shared formula and never typed in.
"Deep cut" means it	Famous places can't pass as secrets.	A place covered by two or more press outlets is relabeled an "open secret."
Right now	Finds what fits the time of day you're in.	Time-of-day labels per place, by Bay Area, San Diego or Portland local time.
Your passport	Stamps, notes, and ratings that are yours.	Stored only on your device. No account needed.
AI as a researcher, not an author of record	Every word was checked by a person before it went live.	Claude drafts cards from web research; drafts stay hidden until reviewed.

How a place gets in

flowchart TB
  OSM["OpenStreetMap<br/>17,506 venues"] --> F["Chain and landmark filter<br/>12,630 independents"]
  F --> D["AI discovery<br/>Claude + web search"]
  D --> S["Selection<br/>dedupe · match to map · balance"]
  S --> R["AI research<br/>drafts each card"]
  R --> H{"Human review"}
  H -- approved --> L["Live guide<br/>253 places"]
  H -- skipped --> X["Left out, with the reason kept"]

Each step is a script in scripts/pipeline/, budget-capped, with a free --estimate mode before anything is spent. The research briefs and rubric live in prompts/. Full detail: How places are found.

Engineering

flowchart LR
  subgraph Build["Build time · Astro"]
    Data[("places.json<br/>reviewed places")] --> Pages["Static pages<br/>front, explore, place, story"]
    Data --> OG["Share images<br/>generated per place"]
  end
  subgraph Edge["Cloudflare"]
    Assets["Static assets"]
    Worker["Worker · forms API"] --> D1[("D1<br/>messages · tips · sign-ups")]
    Worker --> Mail["Email alert<br/>to the owner"]
    Worker --> TS["Turnstile<br/>bot check"]
  end
  Pages --> Assets
  Browser["Browser<br/>passport in localStorage"] --> Assets
  Browser --> Worker

Pages are pre-rendered, so the guide is fast and cheap to run. The only server code handles the contact, suggestion, and sign-up forms. Detail: docs/architecture.

• CI: every pull request and push to main installs from the lockfile, runs the API tests, and runs a full production build.
• Tests: 39 unit tests for the forms API: input validation, the request guards, and the SQL itself, run against the real migrations in an in-memory SQLite. They use Node's built-in test runner, with no test dependencies.
• Forms: every submission must come from the site itself, be JSON under 16 KB, and pass a Cloudflare Turnstile check issued for this hostname. Inputs are validated and length-limited before any query, queries use bound parameters, visitors are identified only by a keyed anonymous hash, and each form has two daily limits: one per visitor and one for the whole site, so a flood from many addresses still stops at a known number.
• Privacy: no accounts, no tracking. Your passport stays on your device. The owner's notification address is a Cloudflare secret, never in the code.
• Browser hardening: a strict content security policy (only the site's own scripts, the few small ones Astro inlines allowed by hash, Turnstile, and the map tiles), HSTS, nosniff, cross-origin window isolation, same-origin framing, a strict referrer policy, and camera, microphone, and payment access disabled. Location is allowed for this site only, for the "near me" sort.
• See SECURITY.md.

npm ci
npm run dev        # http://localhost:4321
npm run build      # static build plus share images

The research pipeline needs an Anthropic key in .env and always offers a free --estimate first:

node --env-file=.env scripts/pipeline/05-enrich.mjs --estimate

src/pages/              Routes: front page, explore, place pages, passport, story, how the score works, forms
src/components/         The passport, ticket cards, neon signs, scores, map sheet, Surprise me deck, header and footer
src/lib/                Scoring formula, place helpers, map style, config
src/data/places.json    The reviewed places that ship
worker/                 Cloudflare Worker: router (index), guards (http, security), validation, D1 store, notifications, tests
migrations/             D1 schema
scripts/pipeline/       The research pipeline, steps 01 to 07
prompts/                Research briefs and the enrichment rubric
data/                   Pipeline inputs, drafts, and review sheets
docs/                   Architecture, data pipeline, decision records, screenshots

How it was built

I designed the product, set the editorial rules (independents only, honest scores, every place reviewed by a person), and directed Claude Code as the engineering team. Claude wrote most of the code and drafted the place research; I reviewed and approved what went live. The working agreement is in AGENTS.md; the reasoning behind the bigger choices is in the decision records.

Stack: Astro 7 · MapLibre GL with OpenFreeMap tiles · Cloudflare Workers, D1, Turnstile, and Email · Anthropic API with web search and Message Batches.

About

Icarian Forge is my personal AI lab, where I build and test what I advise on. The code is © 2026 Dan West, all rights reserved; see LICENSE. Map data © OpenStreetMap contributors, see NOTICE.

Get in touch: LinkedIn · [email protected]