amazon-search

name	amazon-search
description	Run a filtered Amazon US search inside the user's signed-in browser and return a decision table with per-card real delivery dates, Prime-exclusive pricing and brand facets. Use when the answer needs facts that only a signed-in session renders — what arrives today or tomorrow, what it actually costs at checkout, or a search restricted to a trusted brand list. For an anonymous sweep, or when the user has described a need rather than a product, start with amazon-shortlist instead and come here only if the session-only fields are required.
allowed-tools	Read, Write, Edit, Bash(mkdir *), Bash(python3 *), mcp__claude-in-chrome__tabs_context_mcp, mcp__claude-in-chrome__tabs_create_mcp, mcp__claude-in-chrome__navigate, mcp__claude-in-chrome__javascript_tool, mcp__claude-in-chrome__find, mcp__claude-in-chrome__tabs_close_mcp

Amazon search (signed-in)

Scope. This is the browser-driven search. It costs a real browser session and earns that cost only when you need per-card delivery dates, Prime-exclusive prices or the brand facet. For establishing the category and sweeping candidates, amazon-shortlist is cheaper, stateless, and immune to the sticky-filter trap below.

Turn a shopping question into a table the user can act on. Three fields carry the decision, and everything here exists to get them right:

Field	Why it is the one that matters
ASIN	The only unambiguous handle. Titles churn, listings get relisted, search rank moves. Quote it every time.
Real cost	Sticker price is not cost. Shipping fees, basket minima and Prime-exclusive pricing all move it.
Delivery	"Prime" is not a delivery date, and "FREE delivery" is not always free. Say which day, and say what it is conditional on.

Drive the real browser

Use claude-in-chrome against the user's signed-in session. This is not a preference, it is a correctness requirement — the profile's session_dependence records that an anonymous page hides the Prime filter and the curated brand facets entirely and shows a degraded free-shipping stand-in instead. A run without a Prime session answers a different question and looks identical.

Delivery promises are computed against the session's delivery ZIP. Read it (#glow-ingress-line2) and state it with the answer. Do not change it — that is the user's account setting.

Procedure

1. Load the profile

Read profiles/amazon-us.json. Everything volatile lives there: facet IDs, sort keys, selectors, recipes. Do not hardcode any of it in this skill — when Amazon changes, the profile is the single file that gets edited.

Check its verified date. Anything more than a few months old is a hypothesis; re-derive per profiles/README.md rather than trusting it silently.

2. Compose the search

Start from a recipes entry rather than assembling facets from scratch:

• reputable_fast — the default. Prime + tomorrow + sold by Amazon + 4 stars & up.
• same_day — when the user says today. Expect it to be empty for most electronics.
• budget — arbitrary cent ranges work; use the user's real number, never round it to one of the buckets the sidebar offers.
• brand_allowlist — once brand-scrub has built a list.

Two things that will otherwise waste a cycle:

• A facet missing from the sidebar does not mean the token is dead. It means this query and ZIP have no hits for it. Compose from the profile's IDs regardless.
• Facet labels drift even when IDs do not. p_101 renamed itself between two consecutive days because the label names the current cutoff. Never quote a cutoff from the profile — read it live or omit it.

3. Extract

Run extract-results.js from this directory via javascript_tool. It returns parsed rows, never page text. Two constraints are baked into it and must survive any edit:

• Return parsed fields only. The Chrome extension blocks results that look like cookie or query-string data. A row must never contain a raw href, a key=value string or a facet:id token — joining output with = is enough to get the whole call blocked.
• Never pull the page into context. A search page is ~700KB.

For a specific product, run extract-product.js on its /dp/{ASIN} page. Go there when the user is close to deciding, because three facts exist only on that page: who actually sells it, how many competing offers there are and at what price, and how long the delivery promise stays valid.

4. Read the delivery string properly

The promise is one sentence carrying three separate facts. The extractor splits them; this is what the splits mean:

deliverySpeed	Meaning
today / tomorrow	A real promise, against this ZIP, at this hour
dated	A specific day a little further out
estimate-range	Not a promise. "Aug 24 - Sep 1" is an estimate, typically two to three weeks, and with no Prime badge it is the strongest per-card tell of a direct-from-overseas seller

And on cost:

• shipping: "free" — free.
• shipping: "$3.16" — a real fee. Add it to the item price before comparing.
• shippingCondition: "on orders over $25" — the free claim is conditional on basket total. Same-day free delivery carried a $25 minimum, and a $19.98 item's next-day promise carried one too. Calling that free for a single-item order is wrong.
• primeExclusivePrice: true — that price needs a membership. Say so when comparing against a non-Prime retailer.

5. Rank and answer

Filter out sponsored rows before any judgement — they skew heavily to unknown brands buying placement. Apply the profile's trust_rubric, and the brand-scrub allowlist if one exists.

Output shape:

## {query} — Amazon US
Delivering to {ZIP} · read {timestamp} · {resultCount}

| # | Product | ASIN | Price | + shipping | Arrives | Prime |
|---|---------|------|-------|-----------|---------|-------|
| 1 | ...     | B0…  | $19.98| free      | Tomorrow| yes   |

Then, in a line or two: what the filters cost (the resultCount before and after), and anything excluded that the user might have wanted to see. A filter that prunes to nothing is a finding about the category, not a success — say so rather than quietly returning an empty table.

Report rows with no price as "no current offer" rather than dropping them. A missing price is not a cheap price.

What this cannot tell you

Say these plainly rather than improvising around them:

• Review authenticity. Fakespot died 2025-07-01 and nothing replaced it. AI-written reviews read cleanly, so linguistic tells no longer discriminate. Use the structural signals in the profile's trust_rubric — seller identity, whether the brand exists off Amazon — and say the review score is unverified.
• Listing age. "Date First Available" is absent from many electronics listings. When it is missing, say so; do not substitute review count for age.
• Per-card brand. The brand row renders on brand-navigational queries only — 17 of 21 cards on waveshare raspberry pi hat, 0 of 20 on usb c hub. The facet rail is the source of truth. Never build a brand list from cards.
• Price history. No free source. Keepa is the only working programmatic route and it is paid, ~7 tokens per ASIN for a real seller check.

Related

• profiles/amazon-us.json — facet grammar, extraction contract, trust rubric
• brand-scrub — builds the durable brand allowlist this consumes
• amazon-shortlist — category-first, anonymous sweep; start there
• amazon-delivery-check — dates for ASINs you already have, without a search
• reference/search-filters.md — rh= node grammar and the sticky-filter trap
• profiles/README.md — how this profile was derived