find-ultra-quality-food

name	find-ultra-quality-food
description	Audit a set of restaurants/takeaways for two specific health-quality facts about HOW the food is made - (1) are there seed-oil-free options? and (2) are there organic options? - by reading their websites, menus (downloading and OCR'ing PDFs/images), articles and a light pass of reviews with parallel full-strength agents (one per outlet). Each fact gets a yes/no/unknown verdict plus a reliability score (how much to trust it, from the number, recency and type of sources). Produces a ranked, easy-to-scan HTML report. Use when the user wants to find seed-oil-free and/or organic places among an explicit list OR a category + location (e.g. "Indian restaurants in Norwich", "all takeaways in Diss", "restaurants in Paris"). Narrower and cheaper than food-review-review - it extracts two binary facts, it does not cluster every review.
argument-hint	<explicit restaurant list> | "<category> in <place>"

find-ultra-quality-food

Find the places whose food is ultra quality in the specific sense of how it is cooked and sourced — seed-oil-free and/or organic — for a health-conscious diner who avoids industrial seed oils and prefers organic. This is not a taste/authenticity audit (that's food-review-review); it extracts two factual signals and presents them as a clean, ranked report.

The two facts (each a binary verdict + a reliability score)

For each outlet, answer two independent questions. Each gets a verdict of yes / no / unknown and a reliability score (0–100) — how much to trust that verdict.

1. seed_oil_free — is there a way to eat here WITHOUT seed oil? yes if it fries/cooks entirely in non-seed fats OR offers a clear seed-oil-free option a diner can choose (e.g. a beef-dripping fryer alongside the vegetable-oil one — both count as yes, the distinction is deliberately collapsed). no if it cooks in seed/"vegetable" oil with no seed-oil-free option.
2. organic — are there organic options? yes if any organic dishes/ingredients are offered, no if not.

unknown means no evidence was found — it is the honest verdict, never a guess and never a stand-in for no.

What counts as a seed oil

• Seed oils — the ones being avoided: rapeseed/canola, sunflower, safflower, soybean, corn, cottonseed, grapeseed, rice-bran, and generic "vegetable oil" / "blended oil" (in the UK almost always rapeseed or a rapeseed/soy blend — treat as seed oil).
• NOT seed oils — frying only in these ⇒ seed_oil_free: yes: beef dripping, beef tallow, lard, duck/goose fat, butter, ghee/clarified butter, olive oil, coconut oil, avocado oil, palm oil (animal fats or fruit/drupe oils).
• Borderline — peanut/groundnut oil: a legume oil, not a true seed oil, but many avoiders exclude it. Frying only in peanut oil ⇒ yes, but say so in the evidence ("peanut oil — legume, not a seed oil").

Reliability score (0–100)

Reliability is how much to trust the verdict, NOT how good the outlet is. Set it from the number of sources, their recency (old sources are heavily discounted), and their type (official site/menu > article > review), and drop it when sources disagree:

• 90–100 — explicitly stated on the official website/menu (or documented chain policy), current; ideally corroborated.
• 75–90 — multiple independent, recent (≤~2 yrs) reviewers/articles agree.
• 50–65 — a single recent credible source.
• ~10–25 — only old sources (>~3 yrs), e.g. a single review from years ago. (A lone 2018 "fries in dripping" review on a 2026 run ⇒ ~10%.)
• ≤40, and pick the most recent/authoritative verdict — when sources conflict (e.g. an old review says dripping but a recent one says vegetable oil). Recent, authoritative evidence wins the verdict; the disagreement keeps reliability low.
• 0 / unknown — nothing found.

Do not invent reliability from generic priors ("chip shops often use dripping"). A regional prior is not evidence; with no source for this outlet the verdict is unknown.

HARD RULES (do not violate)

1. No evidence ⇒ unknown. Never a default negative. Defaulting a not-found fact to "uses seed oil" / "no organic" is a silent failure and is banned. Never flip a verdict on a hunch.
2. Every non-unknown verdict MUST cite evidence — ≥1 {text, source, date, url?} item with a real quote/paraphrase from a page you fetched, each tagged with its date (so recency can be weighed). No citation ⇒ unknown.
3. Weigh recency and conflicts. Old sources are weak; a recent source overrides a stale one. Reflect this in BOTH the verdict and the reliability (see the scale above).
4. Read real sources with real agents. Each outlet is researched by a full-strength agent that actually fetches the website, finds and downloads + reads the menu (OCR'ing PDF/image menus), searches articles, and does a light review pass. Never substitute a cheaper model and never fabricate facts from the name alone.
5. One agent per outlet, in parallel (see the workflow). Far cheaper than food-review-review.
6. Scale warning. After resolving the outlet list, if it exceeds 60 outlets, print this and continue without waiting (this skill is non-interactive):

   ⚠️  LARGE SET: <N> outlets → <N> research agents, each doing several web fetches.
       This will take a while and use significant tokens. Ctrl-C now to abort.
   

7. Non-interactive. Don't stop to ask questions mid-run. Pick sensible defaults, state them, proceed. (Only exception: the one-time Apify setup, and only if you choose to use it.)

Tools you'll use

• WebSearch / WebFetch (built-in) — the primary engine: official site, menu links, local news / "best beef-dripping chippy" listicles, blog posts, and review snippets.
• Bash + Read — for menus that are PDFs or images: curl -sL "<url>" -o menu.pdf then Read the PDF (or pdftotext menu.pdf -); for an image menu, curl it then Read for OCR. A menu's frying-oil / "all our beef is organic" line is often the single best 100% source.
• Apify MCP (optional, for the review signal) — Google hides reviews behind a consent wall, so to mine reviews for oil/organic mentions you may want Apify (the same server food-review-review uses). Only a light pass is needed: pull reviews and scan for oil/fat/organic mentions with their dates (recency matters). If Apify isn't connected, rely on website/menu/articles + review snippets from WebSearch.
  • Setup: claude mcp add --transport http apify https://mcp.apify.com --header "Authorization: Bearer apify_api_YOUR_TOKEN" (token from https://console.apify.com/settings/integrations), then restart (claude --continue).

Procedure

Step 0 — Resolve the outlet list ($ARGUMENTS)

• Explicit list ("Tom's Fish Bar Mulbarton and Mr Chips Diss") → use those names + localities.
• Category + place ("all takeaways in Diss", "Indian restaurants in Norwich") → discover the outlets first. Prefer the Apify Google Maps places scraper if connected — call-actor with compass/crawler-google-places, input { "searchStringsArray": ["restaurants <place>","takeaway <place>","fish and chips <place>", …per-cuisine…], "maxCrawledPlaces": 100, "language": "en", "countryCode": "gb" } — then read the dataset (get-dataset-items, fields title,categoryName,address,postalCode,city,website,menu,url,placeId). Keep the genuine food outlets in the target locality (filter by city/postcode), drop non-food (barbers, shops) and permanently-closed. If Apify isn't available, enumerate via WebSearch + a local directory and de-duplicate by name.

Record per outlet: clean name, area/town, and website URL if known. Print the resolved list and the HARD RULE 6 warning if > 60. Then proceed.

Step 1 + 2 — Research & verdict per outlet (inside the fan-out agent)

For each outlet the agent:

1. Website — WebSearch "<name> <town>" → WebFetch the official site (about / "our oil" / sustainability / allergen pages).
2. Menu — find it (often a separate PDF/image), download and read it (PDF via Read / pdftotext; image via Read for OCR). Look for the frying medium and "organic" sourcing lines.
3. Articles — WebSearch "<name> <town>" (beef dripping OR "vegetable oil" OR "seed oil" OR organic) and local-press / food-blog coverage.
4. Reviews (light) — Apify review pull or WebSearch review snippets; scan for the oil/fat or organic sourcing, noting each mention's date.

Then set each fact's verdict (yes/no/unknown) from the most recent, authoritative evidence, plus a reliability per the scale, attaching dated evidence items. Nothing found for a fact ⇒ unknown, reliability 0, empty evidence.

Step 3 — Fan out one agent per outlet (Workflow)

Use the Workflow tool. One agent per outlet, full strength (do not pass a cheaper model), agentType: 'general-purpose' so it has WebSearch/WebFetch/Bash/Read (and MCP via ToolSearch). Embed the outlet list as a literal array in the script (the args global has been unreliable), and tell the agents today's date so they can weigh recency.

const OUTLETS = [ { name: "Mr Chips Diss", area: "Diss (IP22 4AB)", website: "…" }, /* …embed all… */ ];
const TODAY = "2026-06-22";  // pass the real current date
const RUBRIC = `…paste "The two facts" + "What counts as a seed oil" + "Reliability score" verbatim…`;
const EV = { type:'array', items:{ type:'object', required:['text','source','date'], properties:{
  text:{type:'string'}, source:{enum:['website','menu','article','review']},
  date:{type:'string'}, url:{type:'string'} } } };
const FACT = (verdicts) => ({ type:'object', required:['verdict','reliability','evidence'], properties:{
  verdict:{enum:verdicts}, reliability:{type:'integer',minimum:0,maximum:100}, evidence:EV } });
const VERDICT_SCHEMA = { type:'object', required:['name','seed_oil_free','organic'], properties:{
  name:{type:'string'}, area:{type:'string'}, website:{type:'string'},
  seed_oil_free: FACT(['yes','no','unknown']),
  organic:       FACT(['yes','no','unknown']),
  take:{type:'string'} } };
phase('Research');
const verdicts = await parallel(OUTLETS.map((o) => () =>
  agent(`${RUBRIC}\n\nToday is ${TODAY}. Research this single outlet and return its verdict.\n` +
        `Name: ${o.name}\nArea: ${o.area}\nKnown website: ${o.website || '(find it)'}\n` +
        `Fetch the website, find+download+read the menu (OCR PDFs/images), search articles, light review pass. ` +
        `Set each fact's verdict from the most recent/authoritative evidence; set reliability from number+recency+type of sources, dropping it on conflict. ` +
        `Cite a dated quote+source for every non-"unknown" verdict. No evidence ⇒ verdict "unknown", reliability 0, evidence []. Never guess.`,
    { label: `research:${o.name}`.slice(0,60), phase:'Research', agentType:'general-purpose', effort:'high', schema:VERDICT_SCHEMA })));
return { restaurants: verdicts.filter(Boolean) };

Step 4 — Assemble combined.json

Take the workflow's restaurants, add a meta block, and write combined.json matching the schema documented at the top of scripts/build_report.py: { meta: {scope, date, sources, notes?}, restaurants: [ {name, area, website?, seed_oil_free, organic, take?} ] }. Use today's date for meta.date. Optionally write a one-line factual take per outlet.

Step 5 — Build the report & publish

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

(<this-skill-dir> = the directory containing this SKILL.md.) Ranking rules (all in the generator):

• seed_oil_free is the primary key; organic breaks ties. Within each fact the order is yes → no → unknown, and within yes the most reliable outlet ranks first (so a strongly-evidenced seed-oil-free option beats a single old review).
• A faded badge = low reliability (badge opacity scales with the score), so a weak yes (e.g. one stale review) is visually obviously weak even though it's still a yes.
• Outlets unknown on BOTH facts are not ranked — they drop into a compressed "no published evidence" table at the foot, keeping the ranked cards scannable.

Publish with the Artifact tool (favicon a food emoji, e.g. 🍟/🥗/🫒). Give the user the URL plus a tight summary: which places are seed-oil-free (and how reliably), which are confirmed seed-oil, which have organic options, and how many came back unknown (and why — e.g. no website/menu online).

Notes

• unknown is a feature, not a gap. A report full of honest unknowns for places with no web presence is correct; quietly marking them "uses seed oil" would be a lie.
• Keep the two facts, the seed-oil definitions, and the reliability scale stable across runs — they are the durable, reusable contract (shared with build_report.py).
• If Apify credit/permissions or a paywalled menu blocks a source, record what you did find and leave the rest unknown; never present an inferred verdict as if evidenced.