Compare Products

Decide what to get (Track A), then find the best price (Track B), or compare it across countries (Track C). Run whichever the ask needs.

• Model not pinned down ("what should I get", "compare these", "is X worth it") → Track A, which hands the winner to Track B for the Before you buy line.
• Exact model known ("where's it cheapest", "is this deal good", "price history", "good time to buy") → Track B directly. Skip Track A's interview.
• Several countries ("SG vs US price", "where in the world is it cheapest", "should I buy this overseas", a trip abroad) → Track C: read references/cross-country.md and follow it.
• "Where can I buy X" → Where to buy, below; then offer Track A or B.

Start from what the owner already has. Before the first question or search, read the host repo's shopping notes (its folder manual, e.g. AGENTS.md, or shopping index says where notes go): brand exclusions and buying preferences; the domain note that lists what is owned and wanted in this category (a wardrobe, skincare or devices note); and recorded sizes. Then:

• Never recommend an excluded brand, even as a budget backup.
• Don't re-recommend something owned; pick what fills a listed gap or pairs with owned items, and say which owned item it pairs with or replaces. Don't ask what the notes already answer.
• Give sizes from the recorded measurements and fit notes, not a generic chart.
• When the owner reports buying something, mark it bought in the wanted list and add it to the owned list in the same session. Tell them where you wrote it.

If the repo keeps none of these notes, ask about owned items and sizes in the Step 1 question.

Track A: decide what to buy

Interview before researching. Generic research disappoints because it searches before it understands the decision. No searching until the Step 3 gate passes. If the ask isn't a purchase recommendation, say so and stop.

1 · Opening question (one message)

Ask for a brain-dump in their own words: what they're buying (category, models in mind), what for (the real job, where/how/how often), what they already know or ruled out and why, budget and where they'll buy (country/currency), what "good" looks like. Close with "...and anything else you think matters." One friendly block, not a form; say a sentence or just the product name is plenty because the next step is tap-to-answer.

2 · Gap analysis (internal)

Score the request against the decision-dimension checklist in REFERENCE.md §1: each dimension known / unknown / not-applicable. Keep only the unknowns that would change the recommendation. Those blind spots are the main value-add.

3 · Clarify the gaps (gate)

Ask the 2–4 highest-leverage unknowns in one round with the structured user-input tool when available, else one compact message. Every question has concrete pre-filled options; never force a long typed answer.

• Most-likely / recommended option first.
• multiSelect: true when several answers can be true at once.
• For a factor the buyer likely didn't know mattered, fold a one-line why it matters into the question.
• Second round only if an answer opens a fork that changes the shortlist.
• Don't ask what they already answered or anything irrelevant to the category.
• "Other" and "just go" are always available; on "just go", state your assumptions and proceed.

4 · Research

Triangulate; never trust a single source.

• Independent hands-on tests, aggregated owner reviews, forum sentiment, spec sheets.
• Hunt failure modes, reliability, common complaints, not just praise.
• Check recency: current model? refresh or seasonal sale imminent?
• Discount affiliate listicles, sponsored placements, astroturf; weight independent testing.
• Marketing is a claim, not evidence. For each claim ask "what spec, test, or mechanism backs this?" and find it. For measurable products lead with independent measurements, standards, teardowns, datasheets; decode buzzwords into the spec that matters (REFERENCE.md §3). Scale depth to stakes: full workup for spec-driven, expensive or health items, light touch for a mug.
• Narrow to 2–4 contenders scored on the buyer's ranked priorities. Never a generic "best overall".

5 · Deliver

Use the template in REFERENCE.md §4: top pick and why it fits them, runner-up ("pick this instead if you weigh X more"), trade-off table across their priorities, what to avoid / common regrets, pre-purchase checks, confidence and what would change the call. Then hand off to Track B.

Track B: find the best price

Lowest legitimate price for a specific model/variant, not a category.

1 · Pin the exact item

Model number, variant, size/colour where price differs, region and currency. Ask only if missing.

2 · Check the local cache first

price-cache.py, in this skill's base directory (run it; --help lists commands), holds prices already gathered, dated, with currency, url and condition. The cache is the host repo's data, not the skill's: it defaults to resources/shopping/price-cache.json under the host repo root ($HOST_REPO, else the Git toplevel of the working directory); --cache or PRICE_CACHE overrides it. Below, PC is python3 <skill dir>/price-cache.py.

• $PC cheapest "<model>" [--currency <CUR>] --available gives an instant answer if we've seen it, and warns when results span currencies.
• $PC stale --days 30 lists what to re-fetch. Re-fetch only stale/missing items; trust the rest.

3 · Sweep current listings, structured sources first

Start from the host repo's venue directory if it keeps one (the shopping notes above), then widen. A CAPTCHA or bot-wall means you're using the hard door; an allowed one almost always exists. In order:

1. Official / aggregator data: retailer or brand API, product feed, RSS, sitemap; aggregators (Google Shopping, PriceSpy, PriceRunner, Idealo) and trackers (Keepa, camelcamelcamel). Stable, no walls, history for free.
2. Live web search across the region's major retailers.
3. web-extract for prices behind JS; it picks Firecrawl or a plain fetch by credit balance.
4. Your own logged-in session (Playwright) where the owner has an account. Legitimate access, not evasion.

Record findings: $PC record <source> <category> "<title>" <lo> <hi> --currency <CUR> [--url <link>] [--condition new|refurb|used]. Always pass --currency; without it prices can't be compared across regions.

On CAPTCHA / bot-detection / paywall, stop and report (web-extract's hard stop) and try another source. Never auto-solve CAPTCHAs or rotate identities/proxies.

4 · Price history

Amazon: camelcamelcamel / Keepa. Elsewhere: PriceSpy / PriceRunner / Idealo / Google Shopping or the category's known tracker (REFERENCE.md §6.4).

5 · Landed cost

Add shipping, taxes/import/duty, subtract valid coupons/cashback. Rank by what you actually pay (REFERENCE.md §6.1).

6 · Verify the winner

Open the listing: price, right model, in stock, ships to region, reputable seller (not grey-market or counterfeit).

Price report

• Cheapest legit option: price, landed cost, retailer, condition, link.
• Runner-up(s): especially a more trusted seller if the cheapest is risky.
• Deal verdict: good now / wait / average against typical and recent low (REFERENCE.md §6.5).
• Watch-outs: fake "was" prices, expired or region-locked coupons, refurb vs new, model-year confusion (REFERENCE.md §6.2–6.3).

Where to buy

List the venues that fit the item and the buyer's region, not a generic web search.

• Start from the host repo's venue directory (the shopping notes above). If it has none, say so and offer to start one; if nothing in it fits, say so and search broadly (Track B step 3).
• Pin the item category and where they will buy (home, online to home, a trip). Ask one short question only if either is unclear.
• Order venues by fit for this item, each with what it is best for and the practical note (location for a shop; shipping, warranty or authenticity caveat online). For authenticity-critical items, prefer brand stores and authorised retailers over marketplaces.
• Give the trade-off in one line ("marketplace cheapest but slow and no warranty; brand store faster with returns"). Respect timing: never suggest an overseas shop for something needed at home next week.
• When the buyer names a new venue worth keeping, add it to the directory. Never invent an address or a detail; leave a TODO:.

Rules

• Honor the buyer's region, currency, and what's actually available to them.
• Cite sources; separate measured fact from reviewer opinion; flag thin evidence.
• Date every price; if unconfirmed, say "verify at checkout".
• Surface refurbished/open-box when the buyer is open to it.
• Legitimate sources only: real retailers, official refurb/used marketplaces, valid public coupons. No leaked keys, dubious grey-market, coupon fraud, or scraping against a site's terms.

Companion skills

• web-extract: web search and page reads, Firecrawl first while its credits last.
• youtube-transcript: mine video reviews for real-world failure modes (Track A).

See references/cross-country.md for Track C and REFERENCE.md: §1 dimension checklist, §2 category blind spots, §3 source rubric, §4 output template, §5 worked example, §6 pricing reference.