A personalized restaurant recommendation engine that lives inside Claude Code. Builds a taste profile through conversation, then searches 30 countries of editorial sources to find places that actually match how you eat.
For people who take restaurants seriously but are tired of TripAdvisor scores and Instagram hype.
Generic restaurant recommendations fail because they optimize for the average visitor. You are not the average visitor. You have specific opinions about cuisine, atmosphere, wine, and what kills an evening.
This skill solves that by maintaining a persistent taste profile — your cuisine hierarchy, anti-patterns, reference restaurants, trusted sources — and applying it every time you search. It searches Reddit diaspora communities in 3 languages, cross-references editorial critics specific to each country, and filters results through your personal calibration. The more you use it, the better it gets.
- Find — searches trusted editorial sources (Reddit, Michelin, Eater, 30 countries of local critics) in up to 3 languages per query
- Filter — every result passes through your taste profile, anti-patterns, and dining history before you see it
- Record — log visits with /10 ratings, building a feedback loop that sharpens future recommendations
- Analyze — periodic pattern analysis reveals what actually drives your ratings and proposes profile updates
- Evolve — the profile grows with every visit, every new city, every 10/10 discovery
Anyone who uses Claude Code and cares about where they eat. The skill is city-agnostic (tested across Madrid, Paris, Berlin, Tokyo, Bangkok, and more) and cuisine-agnostic. It works best if you:
- Have opinions about food beyond "is it good"
- Visit restaurants regularly (couple, friends, or solo)
- Want recommendations backed by editorial sources, not crowd averages
- Are willing to spend 20 minutes on onboarding to calibrate the system
mkdir -p ~/.claude/skills/restaurant
cp SKILL.md ~/.claude/skills/restaurant/SKILL.md
cp local-critics.md ~/.claude/skills/restaurant/local-critics.md/restaurant
Claude detects the skill is unconfigured and starts a guided conversation. Questions come in batches, not one by one:
| Block | Topic | Time | Required? |
|---|---|---|---|
| 1 | City, companions, budget, diet, saved places | ~2 min | Yes |
| 2 | Cuisines, comfort dishes, ingredients | ~3 min | Yes |
| 3 | Food philosophy (product vs concept, returnability) | ~3 min | Recommended |
| 4 | Atmosphere, wine, cocktails, coffee, service | ~3 min | Recommended |
| 5 | Anti-patterns (what kills a restaurant for you) | ~2 min | Recommended |
| 6 | Reference restaurants (5-10 places you love) | ~5 min | Yes |
| 7 | Evening ritual, booking style, trusted sources | ~2 min | Recommended |
Minimum setup: blocks 1 + 2 + 6 (~10 min). Full setup: ~20 min. Skip blocks 3-5, 7 and fill them later.
Onboarding creates all files and folders automatically.
- Claude Code with skill support
- Web search MCP server (Brave Search, Tavily, Exa, or similar) — required, not optional. The skill searches Reddit, Google, and food critic sites on every recommendation.
Find a spot:
/restaurant Madrid sushi
/restaurant Paris wine bar
/restaurant Bangkok street food
/restaurant # Claude asks what you need
Record a visit:
went to Septime in Paris, 9/10
Analyze patterns:
analyze my restaurant preferences
The skill operates in three modes (find, record, analyze), routed by what you type.
Search methodology (find mode):
- Reads your taste profile before every recommendation
- Determines target city's country, selects local editorial sources from a 26-country lookup table
- Searches in up to 3 languages (country + English + cuisine language)
- Runs Reddit queries across city subs, diaspora communities, and cuisine subs
- Builds an anti-recommendation blacklist before selecting candidates
- Cross-checks candidates against your feedback log and saved places
- Outputs 3-5 structured cards sorted by relevance, saves to file
Country-specific intelligence:
- Dominant local platforms override general search in Japan (Tabelog), South Korea (Blue Ribbon Survey), Thailand (Wongnai), Hong Kong (OpenRice), Singapore (Makansutra)
- 30-country editorial database with named critics, publications, and guides (local-critics.md)
- Diaspora search pattern mandatory for ethnic cuisines — the diaspora knows authenticity better than locals
Quality rules:
- Google Maps rating shown but never used as a filter (3.7-rated places can be 9/10)
- Better 2 strong picks than 3 mediocre ones — every recommendation needs a clear story
- No labels, no comparisons to your reference restaurants, no "what to order"
If you prefer to fill the profile by hand instead of onboarding:
mkdir -p ~/Documents/restaurant-data
cp taste-profile-template.md ~/Documents/restaurant-data/taste-profile.md
cp feedback-log-template.md ~/Documents/restaurant-data/feedback-log.mdCreate ~/.claude/skills/restaurant/config.yml:
home_city: Berlin
home_address: "Kastanienallee 7"
data_dir: "/Users/yourname/Documents/restaurant-data"
saved_places_source: none # or google_maps / apple_mapsFill in taste-profile.md — replace HTML comments with your answers.
~/.claude/skills/restaurant/
SKILL.md # Skill definition (search rules, output format, onboarding)
local-critics.md # Editorial food sources by country (30 countries, 750+ entries)
config.yml # Home city, address, settings (created during onboarding)
~/Documents/restaurant-data/ # Your data (path configured in config.yml)
taste-profile.md # Your taste profile
feedback-log.md # Visit log with ratings
cities/ # City recommendation caches
recommendations/ # Saved recommendation outputs
saved-places-data.md # Optional: Google Maps / Apple Maps export
The taste profile is not static. It grows through use:
- Every visit you record recalibrates future recommendations
- After 3-5 visits, Claude suggests filling in skipped profile sections
- After visiting a new city, Claude prompts you to add neighbourhood notes and city quirks
- After a 10/10 visit, Claude suggests adding it to your reference restaurants
- Running "analyze" periodically finds patterns in your ratings and proposes profile updates
- No profile changes happen without your explicit confirmation
| File | Lines | Purpose |
|---|---|---|
| SKILL.md | 350 | Core skill: search rules, output format, onboarding flow |
| local-critics.md | 758 | Editorial food sources for 30 countries (named critics, publications, platforms) |
| taste-profile-template.md | 147 | Empty taste profile with all sections and guidance comments |
| feedback-log-template.md | 72 | Visit log template with rating scale and entry format |
Built on search patterns, taste calibration methods, and output formats refined over months of real-world use across multiple cities and cuisines.
MIT