food-review-review

food-review-review

Find the best places to actually eat in a set, judged on real food quality and authenticity — not on price, speed, or the opinions of people who don't enjoy authentic world food. The output is a ranked, visual report.

Who this is for / what "good" means

The audience is well-travelled, culinarily diverse diners who know authentic world food and are happy to travel or pay for it. Therefore:

• Cost, value-for-money, and delivery speed are NOT quality signals. A place can be pricey and slow and still be the best food. Discount these complaints entirely.
• The goal is to surface genuine quality and authenticity. Praise from someone who clearly knows the cuisine (orders regional/traditional/off-menu dishes and judges them on their own terms) is worth far more than a generic "lovely, best in town".
• Calibrate for palate, not for niceness. A common, legitimately discountable class of negative review comes from diners with an unsophisticated palate who aren't accustomed to authentic world food — e.g. someone who orders chips and gravy from an Indian, or chicken balls from a Chinese, judges a korma as "not spicy", or calls a properly balanced dish "bland" because it isn't drowned in sweet sauce. Their complaints reflect their expectations, not the kitchen. Treat these as noise. Keep the language neutral and descriptive — never insulting.
• Conversely, credible food faults are gold and must be kept: undercooking, illness, foreign objects, hygiene, or a discerning diner explaining a genuine recipe/technique failure.

The taxonomy (six judgements + one bucket)

Every review with text is placed in exactly ONE cluster. These are cuisine-agnostic; the examples just illustrate.

Cluster	Meaning	Verdict
non_food_complaint	Substance is ONLY non-food: delivery time, lateness, waiting, price/value, wrong or missing order, packaging, parking, staff rudeness/service — no judgement of how the food tasted.	Discount
unsophisticated_complaint	Negative about the FOOD but from a palate unaccustomed to authentic world cuisine: judges by anglicised/Westernised staples (chips, chicken balls, "curry sauce"), calls properly cooked food "bland/no flavour/not spicy enough", expects MSG or heavy sweet sauce, objects to authentic texture/spicing, vague "greasy" with no specifics.	Discount
mixed_or_other	Genuinely mixed (good food + a non-food gripe) or impossible to place.	Set aside
credible_food_complaint	Negative about the food citing a concrete, legitimate defect OR from a clearly discerning diner: raw/undercooked, made-ill, foreign object, hygiene rating, "the [regional dish] should be made with X not Y", "every dish tastes the same", a proper assessment of a technique.	Keep — real signal
generic_praise	Positive but no authenticity signal: "lovely", "best in town", "hot and tasty", great portions/value/service, praises anglicised staples.	Positive, low weight
authentic_praise	Positive AND food-literate: names traditional/regional/off-menu dishes and judges them on their own terms; recognises proper recipes vs cheap shortcuts; orders the things a knowledgeable diner of that cuisine would.	Trust — strong signal
rating_only	A star rating with NO written text. No content to judge.	Separate bucket (counted for volume only)

The ranking score rewards authentic_praise (×2) and generic_praise (×1), penalises credible_food_complaint (×2) and lightly unsophisticated_complaint (×0.5); non_food, rating_only, and mixed are excluded. (The bundled report computes and displays this.)

HARD RULES (do not violate)

1. Read every review with a real agent. Each review with text MUST be classified by a full-strength model agent that has actually read its text. NEVER sample, NEVER substitute a cheaper/smaller model for the classification, and NEVER replace the agents with an embedding/keyword/ML clustering algorithm. The whole point is human-grade judgement at scale. (Use many agents in parallel — see the workflow below.) This is expensive in tokens by design, and the user accepts that.
2. Token-cost warning. After counting the total reviews to be classified, if it exceeds 500, print this banner prominently and then continue without waiting (this skill is non-interactive — the user can Ctrl-C to abort):

   ⚠️  THIS WORKFLOW WILL BE VERY TOKEN EXPENSIVE DUE TO LARGE NUMBER OF REVIEWS
       (<N> reviews → roughly <N/12> classifier agents). Cancel now if you don't want this.
   

3. Non-interactive. Do not stop to ask the user questions mid-run. Make sensible defaults, state them, and proceed. (Only exception: the one-time Apify setup, if it isn't connected.)
4. Both platforms, in full. Harvest all Google and all TripAdvisor reviews. Don't stop at page 1; don't rely on aggregator summaries.

Prerequisite: Apify (one-time setup)

Reviews are harvested with Apify scrapers via its MCP server. Google Maps hides reviews behind a cookie-consent wall that defeats plain fetching, and TripAdvisor paginates awkwardly — Apify solves both and returns clean structured data with real timestamps.

Check if it's already connected: run claude mcp list (look for apify ... ✔ Connected) or search for the tools (mcp__apify__call-actor, mcp__apify__get-dataset-items). If connected, skip ahead.

If not connected, walk the user through this (it's quick and cheap — free tier covers normal use; the Google reviews scraper is ~$0.0006/review):

1. Create a free account at https://apify.com.
2. Copy the API token from https://console.apify.com/settings/integrations (apify_api_…).
3. Add the MCP server (token in the header avoids the OAuth dialog):

   claude mcp add --transport http apify https://mcp.apify.com --header "Authorization: Bearer apify_api_YOUR_TOKEN"
   

   (Alternatively claude mcp add --transport http apify https://mcp.apify.com then /mcp → select apify → complete the browser OAuth.)
4. Restart Claude Code, resuming the session: claude --continue. A running session does not hot-load a newly added MCP server — this restart is required, or the tools won't appear.
5. Verify with claude mcp list → apify … ✔ Connected, then re-invoke this skill.

Procedure

Step 0 — Interpret the argument ($ARGUMENTS)

• Explicit list ("Golden Dragon, Happy Palace and Imperial Wok in Diss") → use those names + their locality.
• Category + place ("Indian restaurants in Norwich") → first discover the outlets. Use the Apify Google Maps places scraper to enumerate candidates: mcp__apify__call-actor with actor compass/crawler-google-places, input { "searchStringsArray": ["Indian restaurant Norwich"], "maxCrawledPlaces": 30, "language": "en" }, then read the dataset (mcp__apify__get-dataset-items) and keep the genuine matches of that cuisine (use title, categoryName, reviewsCount, url/placeId, totalScore). Present the resolved list (name, area, Google review count) and proceed.

Record for each outlet: a clean name, its Google Maps place URL or placeId, and (if found) its TripAdvisor review-page URL.

Step 1 — Harvest ALL reviews (Apify)

Always mcp__apify__fetch-actor-details (output {inputSchema:true,pricing:true}) before running an actor, then call-actor with waitSecs:0 (fire-and-forget) and poll with get-actor-run (waitSecs:45) until SUCCEEDED.

Google — actor compass/google-maps-reviews-scraper (verify via search-actors "Google Maps reviews"):

{ "startUrls": [{"url":"https://www.google.com/maps/search/<Name+Full+Address>"}, ...],
  "maxReviews": 1000, "reviewsSort": "newest", "language": "en",
  "reviewsOrigin": "google", "personalData": true }

reviewsOrigin:"google" keeps Google-native reviews only (so they don't duplicate TripAdvisor). Pass one startUrl per outlet; a /maps/search/ URL with the full name + street + town + postcode resolves to the right place. After it finishes, verify each place via a compact get-dataset-items projection (title,address,city,postalCode,reviewsCount,totalScore) — confirm the address/town matches and that scraped count ≈ reviewsCount.

TripAdvisor — find the actor with search-actors "Tripadvisor reviews" (e.g. maxcopell/tripadvisor-reviews), read its input schema, and run it with each outlet's TripAdvisor URL (or let it search by name+location), pulling all reviews. If no TripAdvisor actor is available, fall back to WebFetch on the TripAdvisor review pages, paginating with the -orN- offset in the URL (e.g. …-Reviews-or10-…, -or20-, …) and de-duplicating.

Get the data onto disk. get-dataset-items results that exceed the token cap are written to a file automatically (the tool tells you the path) — that's the easiest route. Otherwise project fields and page through. You need at least: review text, stars/rating, publish date (for the year), reviewer name, and which outlet + source it belongs to.

Step 2 — Normalise, count, batch (Python)

Write a small Python script (in the scratchpad) that loads the harvested files and produces, for every review, a record: {id, restaurant, reviewer, rating(1-5), year(int), title, text, source}. Map each Google place title/placeId to its outlet name; derive year from the publish timestamp (convert any relative dates against today). De-duplicate by review id / (name+text).

• Split reviews with text into batch files of ~12 (batch_000.json …), each an array of {id, restaurant, rating, title, text}.
• Reviews without text → a rating_only.json list (classified deterministically, not by an agent).

Print the total and the Step-1 §HARD RULE 2 warning if > 500.

Step 3 — Classify EVERY text review with parallel agents (Workflow)

Use the Workflow tool. One agent per batch file: it Reads the file and classifies each review against the rubric, returning structured output. Use the session model at high effort — do not pass a cheaper model. (Implementation note: pass the batch-file paths by embedding them as a literal array in the workflow script; the args global has been unreliable.)

Give each classifier agent the exact rubric from "The taxonomy" above, and this schema per review: {id, cluster (the 6 content clusters), sentiment (positive|negative|mixed), authenticity_signal (literate|anglicised|neutral), dishes_mentioned (string[]), rationale}.

Sketch:

const paths = [ "/abs/batch_000.json", /* …embed every batch path here… */ ];
const RUBRIC = `…the taxonomy text, verbatim…`;
const CLASSIFY_SCHEMA = { /* object with classifications: [{id,cluster,sentiment,authenticity_signal,dishes_mentioned,rationale}] */ };
phase('Classify');
const results = await parallel(paths.map((p,i) => () =>
  agent(`${RUBRIC}\nRead the JSON file at:\n${p}\nIt is an array of reviews (id, restaurant, rating, title, text). Classify EVERY one, keyed by id.`,
    { label:`classify:${i}`, phase:'Classify', agentType:'general-purpose', effort:'high', schema:CLASSIFY_SCHEMA })
    .then(r => (r && r.classifications) || []) ));
return { classifications: results.filter(Boolean).flat() };

For categories with many outlets you may have a separate harvest workflow too, but classification is the part that MUST fan out one-agent-reads-each.

Step 4 — Merge & aggregate (Python)

Join the classifications back to the review records by id. Add the rating_only reviews (cluster rating_only; sentiment from stars: ≥4 positive, 3 mixed, ≤2 negative). Then compute:

• counts.byCluster, counts.byRestaurant (name → cluster → n), counts.bySource
• summary: dismissible (= non_food + unsophisticated), credible_negative, positive (= authentic + generic), rating_only, mixed
• histogram: years (sorted), overall (year → cluster → n), byRestaurant (name → year → cluster → n)
• restaurants[]: {name, score_label (e.g. "Google 4.5★"), total, badge, take} — you write the one-line badge and 1–2-sentence take per outlet from its cluster mix + notable quotes.
• quotes[]: ~6–8 hand-picked verbatim examples per pile (authentic_praise, credible_food_complaint, and the discountable non_food/unsophisticated), each {reviewer, restaurant, year, rating, text, cluster, source}. Truncate text to ~320 chars.
• meta: {scope (e.g. "Indian restaurants in Norwich"), total, date, sources (labels)}

Write it all to combined.json in the schema documented at the top of scripts/build_report.py.

Step 5 — Build the report & publish

Run the bundled generator (vertical layout + ranked leaderboard, scales to any number of outlets). It ships alongside this skill at scripts/build_report.py — resolve that path relative to this SKILL.md's own directory (the skill's plugin install location):

python3 <this-skill-dir>/scripts/build_report.py <combined.json> <out.html>

Then publish with the Artifact tool (favicon a food emoji, e.g. 🍜/🥡/🍛). Give the user the URL plus a tight prose summary: the ranked top picks, the totals (dismissible / credible / positive / star-only), and any surprises (e.g. an outlet that looks great on stars but hides a food-safety pattern). Always restate the caveats: classification is automated ("~this many"), and star-only ratings are bucketed separately.

Notes

• The report's design is a warm "world-cuisine menu" identity (lacquer red, gold, ink, jade). It is intentionally self-contained (inline CSS/JS) so the Artifact CSP is satisfied.
• Keep the cluster names and definitions stable across runs — they're the durable, reusable part.
• If Apify credit/permissions block a scrape, report exactly what failed and what was still obtained; never silently fall back to a thin aggregator sample and present it as complete.