amazon-au-research

name	amazon-au-research
description	Research products on amazon.com.au (Australia) and return options with title-verified links, indicative AUD prices, and stock status. Use when asked to research a product, "find me options for X", compare prices, or for help shopping or buying something on Amazon. Australian marketplace only — this skill does not cover amazon.com or other marketplaces.
allowed-tools	Bash(curl *), Bash(python3 *), Bash(sleep *), Read
user-invocable	true

Amazon AU product research

Produce a short, verified shortlist of products on amazon.com.au. Every link you hand back must be one you resolved to a real AU product title yourself.

Why this skill exists

Four failure modes make the obvious approaches produce confidently wrong answers:

1. Web search cannot source AU links. Search engines return amazon.com results almost exclusively, even when the query names amazon.com.au explicitly. Do not use web search to find product links. Scrape the AU search page instead.
2. ASINs are not portable between marketplaces. A /dp/<ASIN> that works on amazon.com usually does not exist on amazon.com.au. Measured: of 3 US ASINs, 1 existed on AU — and it was a different variant of the product. Never translate a .com link into a .com.au link by swapping the domain.
3. WebFetch is 503'd by Amazon. Use curl with a browser User-Agent and an Accept-Language: en-AU,en header.
4. HTTP 200 does not mean the product exists. Amazon serves 200 for placeholder pages. The <title> is the only reliable existence check.

Workflow

1. Find real ASINs from the AU search page

UA="Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120 Safari/537.36"
curl -s --compressed -A "$UA" -H "Accept-Language: en-AU,en" --max-time 30 \
  "https://www.amazon.com.au/s?k=tech+pouch" \
  | python3 -c 'import sys,re; h=sys.stdin.read(); seen=[]; [seen.append(a) for a in re.findall(r"data-asin=\"([A-Z0-9]{10})\"",h) if a not in seen]; print("\n".join(seen[:15]))'

Keep --compressed: Amazon gzips regardless, so without it Python dies on UnicodeDecodeError: invalid start byte at position 1.

URL-encode the query (spaces as +). Run two or three differently-phrased searches when the brief is loose — one query's top hits are not a survey.

2. Resolve each ASIN to a title, price, and stock

resolve_asin.py sits next to this file. Pipe a product page into it:

UA="Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120 Safari/537.36"
SKILL_DIR="$HOME/.agents/skills/amazon-au-research"   # or ~/.claude/skills/amazon-au-research
for A in B0FCVLH4XS B0DPQPGXBK; do
  curl -s --compressed -A "$UA" -H "Accept-Language: en-AU,en" --max-time 30 "https://www.amazon.com.au/dp/$A" \
    | python3 "$SKILL_DIR/resolve_asin.py" "$A"
  sleep 2
done

Output is TSV: ASIN <tab> price <tab> stock <tab> title, e.g.

B0FCVLH4XS	$99.00	Only 4 left in stock.	OrbitKey 2-in-1 Tech Pouch (Olive)
B0CZ4ZZZZZ	INVALID	Page Not Found

Discard every INVALID row. Never include one in the output, and never report a link you did not resolve. Keep the sleep 2 — this is human browsing volume, not a crawl.

Amazon throttles bursts, and a throttled response yields a blank or partial row rather than an error. If a row comes back empty or a batch that previously worked starts failing, wait and retry that ASIN before concluding it is invalid — sleep 3 between requests clears it. A single INVALID amid otherwise good rows is trustworthy; a run of them means you are being throttled, not that the products vanished.

If you are writing throwaway extraction code rather than using the script, note that Python f-strings cannot contain backslash-escaped quotes. Use %-formatting or .format(); an f-string with \" inside is a syntax error.

3. Judge the candidates

Resolve more ASINs than you need (8–12 for a shortlist of 3) and pick on fit, not on search rank. Search order reflects Amazon's advertising and sales signals, not the user's brief. Read the titles for the attributes the brief actually asked about — size, material, capacity, compatibility.

Verify hard constraints against the full page, not the resolved title

When the brief has a pass/fail requirement — a bundled accessory, a port, a capacity, a compatibility — the TSV row is not enough to clear it. resolve_asin.py truncates titles at ~100 characters, and AU listings put the disqualifying clause at the end of a long title, where it gets cut. AU listings also sell stripped variants under a name that otherwise reads identical to the full product. Fetch the page and grep the feature bullets and the manufacturer copy for the constraint before shortlisting:

curl -s --compressed -A "$UA" -H "Accept-Language: en-AU,en" --max-time 30 "https://www.amazon.com.au/dp/$A" \
| python3 -c '
import sys,re,html
h=re.sub(r"\s+"," ",html.unescape(re.sub(r"<[^>]+>"," ",sys.stdin.read())))
for m in re.finditer(r"(?i).{80}(logi bolt|usb receiver).{80}",h): print(" *",m.group(0).strip()[:190])
' | sort -u | head -6

Measured, Jul 2026: the "MX Master 3S – Bluetooth Edition" (B0FNV6GP6K) declares "Logi Bolt receiver not included" — in its bullets and in the tail of its own title, past the truncation point. AU buyers of the MX Anywhere 3S (B0C6K2DH73) report the receiver missing there too, with nothing in the listing saying so. The standard 3S (B0B11LJ69K) confirms it ships. Three near-identical resolved titles, opposite answers to the actual brief.

Reviews are usable as corroboration for this and nothing else: a constraint denied by several buyers is worth flagging even when the official copy is silent.

Output contract

Return N options (default 3), each with:

• Product title.
• A /dp/<ASIN> link on www.amazon.com.au that you title-verified in step 2.
• Indicative price in AUD.
• Stock status.
• One line on why it fits the brief.

Then:

• An explicit recommendation — name one, say why.
• Honest flagging of what the products do not solve. If nothing found meets part of the brief, say so plainly rather than stretching a product to fit.

Prices are approximate

Several dollar figures appear in a product page's HTML (list price, deal price, variant prices, subscription prices). The scraped figure is a reasonable indication, not a quote. Always label prices as indicative and tell the user to confirm on the page — especially before recommending on price grounds.

Caveats to state, not bury

• Scraping is against Amazon's ToS. Fine at human browsing volume; do not build anything that hammers it. Expect selectors to break — if extraction returns empty, the markup changed.
• The official APIs are not a viable alternative. PA-API 5.0 was deprecated 15 May 2026 and stopped accepting new signups; its replacement, the Creators API, requires an active Associates account with roughly 10 qualifying sales per rolling 30 days. If price history matters, Keepa (paid) is the only decent option.