Jonathon's AI Wiki

Claude Shopping Assistant — A Project + Cowork Recipe for Buying Quality (Nicole Ruiz)

5 min read

Source: raw/She_built_a_Claude_shopping_assistant_to_stop_buying_cheap_junk.md — How I AI (Claire Vo), guest Nicole Ruiz (@nicolewilliams30; Substack on technology in the household). YouTube: https://www.youtube.com/watch?v=OOPganyUinE. Fetched 2026-06-08.

A “How I AI” walkthrough of a personal-productivity build: a claude.ai Project that acts as a vetting-and-research shopping assistant for high-quality, long-lasting goods, paired with Claude Cowork (and mobile Dispatch) wired to Gmail to automate product returns. The reusable artifact is the prompt scaffold — a stored set of purchasing criteria plus a fixed output format — that turns a recurring, judgment-heavy chore into a single short query. The framing: move the decision-making upstream into reusable instructions so the recurring task (“is this well-made, returnable, natural-fiber, in-budget?”) costs near-zero mental overhead.

Key Takeaways

  • The Project holds the criteria, not the query. A dedicated claude.ai Project stores (a) a trusted-vendor list imported from an Apple Notes dump, (b) the criteria for why a vendor qualifies (decades in business, makership reputation, made-to-last, honors returns), and (c) fallback rules for finding new preferred vendors. Keeping this in a Project (vs. the default chat) avoids overfitting Claude’s general queries to these purchasing instructions and keeps purchase memory separate.
  • A fixed output format is the load-bearing part of the prompt. Each result must surface: product name, a photo, price, materials (flagged because “plastic” is often buried low), care/maintenance notes (e.g. handwash-only), a purchasable link, and a one-line trustworthiness note on the brand’s history. The brand-history line is the highest-value field — it surfaces signals like “brand was acquired two years ago and reviews collapsed; don’t buy.”
  • Explicit anti-patterns in the instructions. The prompt tells Claude to avoid trendy DTC brands that over-spend on ads and under-invest in quality, to discount AI-sounding reviews, and to flag likely drop-shippers — encoding modern “shopping-internet pitfalls” as filters.
  • It levels the field for badly-built sites. Heritage manufacturers often have the worst, least-searchable websites; the assistant reads them anyway, so a “no UX is the best UX” pass-through gets the user to the product without browsing.
  • Returns are a second surface: Cowork + Gmail. For a failed product, she switches to Claude Cowork (mobile via Dispatch, often dictated with Whisper Flow while hands-free): photograph the item, ask Claude to find the receipt in Gmail (PayPal or merchant) with item number/order details, then draft a refund email to customer service. The angle is “this was supposed to last,” not “within the 30-day window” — holding the brand to its quality promise.
  • The loop improves with feedback. Telling the Project what she actually bought refines future recommendations and surfaces new vendors; sizing (size-guide normalization) is the part she has not yet solved, so clothing is still human-in-the-loop.
  • Not yet fully autonomous. She wants to reach standing/recurring orders (Costco-style) but says nothing is regular enough yet — the project still needs more iteration before it buys unattended.
  • Prompting-under-failure tip. When Claude misses, ask “why did you do that / where did we go wrong?” before re-instructing — get the model’s reasoning, then add the corrected criteria as a new guideline (treat it like managing an employee).

Try It

  1. Create a dedicated claude.ai Project for purchasing. In its custom instructions, paste your own trusted-vendor list and the criteria that earn a vendor a spot (longevity, makership, returns honored, materials).
  2. Lock an output format in the instructions: name, photo, price, materials, care notes, purchase link, one-line brand-history/trust note. Add anti-patterns: discount AI-sounding reviews, flag drop-shippers, deprioritize ad-heavy DTC brands.
  3. Query short: “Help me find a [item].” The Project supplies all context, so the prompt stays one line.
  4. Add a budget/gift-card angle when relevant: “I have $30 for [brand]; what fits my criteria?” or set a price band.
  5. Vet a new brand before buying: “What’s your analysis of this brand — are they legitimate?” to pull acquisition history, review trends, and quality signals.
  6. Automate returns in Cowork (mobile via Dispatch): photograph the item, ask Claude to find the receipt in Gmail and draft a refund request citing the quality shortfall. Review, then send.
  7. Close the loop: tell the Project what you actually bought so recommendations sharpen over time.

Open Questions

  • Sizing normalization is unsolved — she explicitly has not gotten Claude to reliably reconcile per-brand size guides against general international sizing, so clothing buying stays manual. ^[inferred]
  • Autonomous purchasing not yet reached — she references another operator running Instacart/Amazon buy-flows but has not implemented standing orders herself; the threshold for “regular enough to automate” is undefined.

Graph View

Backlinks