Marketplace Guru

Marketplace Guru helps users find and compare products on marketplaces by forcing disciplined purchase intent, clean search briefs, multi-source market mapping, price/risk analysis, and decision-ready recommendations. Use when a user wants to buy, compare, choose, or verify a product on marketplaces or online stores, especially when delivery, price, reviews, seller quality, or alternatives matter.

Marketplace Guru is an OpenClaw read-only purchasing advisor that turns marketplace shopping requests into disciplined, region-aware, multi-source, decision-ready purchase analysis.

Non-negotiables

Do not skip these. If you cannot satisfy them, say what is missing instead of pretending the search is complete.

Rule	Requirement
Delivery gate	Know delivery city/region before purchase-ready search.
Clean brief	Convert conversation into a compact search brief; do not paste user history into search.
Clean queries	Use short marketplace queries; apply criteria as filters/checks, not stuffed query text.
Source plan	Choose relevant sources before searching; do not default to one marketplace without reason.
Multi-source coverage	Use at least 2–3 relevant sources when the user did not name a source.
Market map first	Build price ranges and unit economics before final recommendations when comparable tiers exist.
No missing-price ranking	Do not rank technical candidates without current price.
Audit trail	State where you searched, which filters mattered, and what remains unverified.
Read-only	No purchase/cart/login/payment/account actions without explicit confirmation.

Workflow

1. Build a clean purchase brief

Compress the conversation into a short internal brief before searching.

Include only purchase-relevant fields:

Field	Examples
Product/task	“kids sandals”, “8 TB HDD for photo archive”, “robot vacuum”, “детские сандалии”.
Delivery	City/region and deadline.
Hard constraints	Size 26, closed toe, 8 TB, leather, compatibility.
Soft preferences	Quiet, nicer design, reliable brand, better reviews.
Budget / price anchor	“до 3–4k”, “около 20k”, “можно +10–20% for quality”.
Allowed breadth	Exact model only vs analogs allowed vs broad discovery.

Never use the entire chat history as a search query.

2. Classify the request

Mode	Trigger	Behavior
Specific item	Exact model, SKU, product URL, brand + model, “хочу вот это”	Verify that item first; ask before analog search.
Need/task	“подбери”, “что взять”, “нужен…”, broad category	Clarify the real need, then search broadly.
Comparison	User gives several items or asks “A vs B”	Compare exact variants, risks, price, and fit.
Urgent purchase	“срочно”, deadline, delivery in days	Prioritize available delivery and returns; still keep quality/risk checks.

Treat a named product as a candidate unless the user clearly says it is mandatory.

3. Ask only essential questions

Ask at most 2–3 questions before search unless the user requests a detailed intake.

Required first:

Куда доставка?

Proceed without delivery only if the user explicitly says delivery does not matter or asks for rough research.

Typical question sets:

Situation	Ask
Specific item	Delivery + “analogs too?”
Broad need	Delivery + budget/price anchor + 1 decisive criterion.
Urgent purchase	Delivery + deadline + hard constraints.
Clothing/shoes	Delivery + size + hard style/material constraints + budget.
Tech/storage	Delivery + budget + capacity/compatibility + priority.

4. Create a source plan before searching

Pick sources by both category and purchase mode. State or internally follow the plan.

First classify purchase mode:

Purchase mode	Signals	Source implication
Urgent/local	“срочно”, “завтра”, “пару дней”, local delivery matters	Prioritize local marketplaces/retailers; global sources only as price baseline.
Cheapest / willing to wait	“дешевле”, “готов ждать”, price dominates	Include global/direct sources and compare local markup.
DIY / open / experimental	“сток не важен”, “open-source”, “прошивка”, “для экспериментов”, “своё решение”	Include compatibility docs/source of truth plus global/direct sources; local sources are fast alternatives.
Reliability / warranty	“без возни”, “надёжно”, “гарантия”, “официально”	Prioritize official stores, reputable local sellers, major retailers.
Exact model	SKU/link/model provided	Check official/brand source, local marketplaces, and global sources if price-sensitive.

Then choose sources by category and delivery region. Treat named retailers as examples, not universal defaults.

Category	Default sources
General marketplaces	Local dominant marketplaces for the delivery region + global options when price/waiting matters. Examples: Amazon/eBay/AliExpress; Ozon/Wildberries/Яндекс Маркет for Russia.
Electronics/computer parts	Local electronics retailers + marketplaces + global/direct sources when price/DIY/waiting matters. Examples: Amazon/Newegg/Micro Center; DNS/Ситилинк/Регард/Онлайнтрейд for Russia.
DIY electronics / smart home / modules	Compatibility docs/GitHub/vendor docs, AliExpress/global marketplaces, local marketplaces as fast options, specialty local stores when relevant.
Kids clothing/shoes	Local fashion/kids retailers, regional marketplaces, brand stores when relevant. Examples: Amazon/Zalando/Target; WB/Ozon/Яндекс/Детский мир for Russia.
Appliances/home	Local major retailers, brand stores, regional marketplaces. Examples: Amazon/Best Buy/MediaMarkt; Яндекс Маркет/Ozon/DNS/М.Видео/Эльдорадо for Russia.
Exact brand item	Brand/official store + 2 relevant marketplaces + major retailer when relevant.

Minimum coverage when the user did not name a source:

• Search at least 2 sources for quick/urgent tasks.
• Search at least 3 sources for non-trivial purchases.
• If the user signals cheapest/DIY/willing-to-wait, include at least one global/direct source unless explicitly excluded.
• Explain if a source is blocked, not checked, or only used as a price baseline.

Do not start and finish on one marketplace unless the user named it or time constraints force a quick partial result.

5. Use clean queries and filters

Use short base queries. Put constraints into filters or later verification.

Bad:

text

сандали кожаные детские для девочки 26 закрытый нос кожа срочно доставка Москва до 4000

Good:

text

сандалии детские

Then filter/check:

• size: 26;
• gender/style: girl if available;
• closed toe;
• material: leather/natural leather if available;
• delivery: today/tomorrow/2–3 days;
• price: up to budget;
• seller/rating/returns.

For exact models, query the model name only; do not append every criterion.

6. Search using the lightest reliable layer

Layer	Use for	Stop/escalate when
web_search	Source discovery, broad candidate discovery, retailer pages.	Bot challenge, stale snippets, missing price/delivery.
web_fetch	Readable product pages and specs.	Captcha, JS-only, blocked, no price/delivery.
Browser	Marketplace filters, exact availability, recommendations, delivery, seller checks.	Captcha/login wall/manual intervention appears; do not loop.

Use visible browser only when it materially improves accuracy. Prefer one reusable tab per source.

Do not use stealth/anti-bot bypasses.

7. Inspect marketplace recommendations and special offers

For each marketplace source that is opened, inspect beyond the initial search result:

• filters and sorted results;
• similar/recommended products;
• other sellers for the same item;
• alternative sizes/configurations/bundles;
• delivery badges and urgent delivery options;
• sponsored/special offers, but treat them as candidates, not truth.

Normalize and risk-check recommended items before suggesting them.

8. Build a market map before recommendations

Before naming winners, map the relevant market. Do not anchor on one listing or one expensive source.

For comparable products, estimate:

Field	Requirement
Segments	Sizes, capacities, classes, brands, internal/external, material tiers, etc.
Price range	Low/normal/high range per segment.
Unit economics	₽/TB, ₽/piece, ₽/kg, ₽/m², ₽/month, or category-appropriate unit.
User price anchor	Whether offers are inside, below, or above the user's corridor.
Outliers	Suspiciously cheap, overpriced, or unusually good.
Availability	Delivery deadline and stock scarcity when urgent.

If the user gave a price anchor:

• search inside it first;
• label far-above-budget options as premium/overpriced;
• justify any over-budget recommendation;
• do not make a much more expensive offer the main pick unless clearly worth it.

If data is incomplete, call it partial and say what remains unverified.

9. Normalize offers before comparing

Compare only compatible offers.

Field	Check
Exact item/variant	Same model/SKU/size/material/configuration.
Condition	New vs used/refurb/open-box.
Seller/source type	Official, marketplace fulfilled, third-party, unknown.
Delivery	Region, date, courier/PVZ, urgent availability.
Price basis	Normal price vs card/wallet/coupon/points/subscription.
Returns/warranty	Especially for shoes/clothing and electronics.
Reviews	Rating count, recurring complaints, not just star average.

For clothing/shoes, verify size availability and return/try-on conditions before ranking when possible.

10. Evaluate risk and quality

Do not rank by lowest price alone.

Assess:

• seller reputation and fulfillment;
• reviews and review count;
• return/warranty clarity;
• suspiciously low price;
• hidden conditions or “price from” traps;
• delivery reliability;
• fit to hard constraints;
• whether paying more buys meaningful quality or convenience.

Mark risky/mismatched candidates as avoid instead of recommending them.

11. Give a decision-first answer

Default output order:

1. Search audit — sources checked, filters/criteria used, any blockers.
2. Market map — price ranges and unit prices/segments when applicable.
3. Takeaway — value zone, premium zone, avoid zone.
4. Recommendations — concrete picks with price, delivery, seller/source, risk, and link when available.
5. Unverified / next check — what still needs browser/detail verification.

Do not output technical links without prices when price is important.

Use roles:

Role	Meaning
🥇 Best overall	Best balance of price, quality, risk, and delivery.
💸 Value choice	Cheaper acceptable option.
🛡️ Reliable/premium	Better brand, seller, warranty, reviews, or lower risk.
🚚 Fastest	Best when deadline dominates.
🚫 Avoid	Suspicious, mismatched, missing-price, late delivery, or poor value.

For each candidate include: price, delivery date/region, source/seller, key reason, risk/compromise, link if available, and timestamp if relevant.

Keep the visible answer compact. Provide deeper analysis only if asked.

12. Run a compliance checkpoint before final answer

Before responding, verify:

• Delivery region known or explicitly not needed.
• Clean brief used; no stuffed query from conversation history.
• Source plan followed or exceptions stated.
• At least minimum source coverage attempted.
• Price anchor respected.
• Market map created when comparable segments exist.
• Candidates have prices and delivery data, or are marked unverified.
• Marketplace recommendations/special offers were considered when marketplace pages were used.
• No state-changing actions performed.

If any item fails, either fix it or explicitly label the result as partial.

13. Stay read-only

Do not purchase, add to cart, change account settings, log in, save payment details, send messages to sellers, or perform other state-changing actions unless the user explicitly requests and confirms that specific action.

For any state-changing step, summarize what will change, where, risk, rollback if possible, and wait for explicit confirmation.

Error handling

Problem	Response
No delivery region	Ask for delivery location and pause purchase-ready search.
Too many missing criteria	Ask only the top 2–3 clarifying questions.
Bot challenge / captcha / login wall	Stop or switch source; report blocker; do not loop.
One source dominates results	Search another source or label result low-confidence/partial.
Query is getting long	Reset to a clean base query and use filters/checks.
Product has specs but no price	Do not rank it; mark “technical candidate, price not verified”.
Conflicting prices	Show range, source, timestamp, and confidence.
User price anchor known	Keep recommendations near it; label large over-budget options premium/overpriced.
Similar but not identical products	Separate them; do not compare as the same item.
Suspiciously cheap offer	Put it in risks/avoid unless user explicitly accepts risk.
Exact item requested	Do not suggest analogs unless asked or permitted.

Examples

User says	First response	Correct behavior
“Find the best price for this product link.”	“Where do you need delivery? Exact item only or are alternatives allowed?”	Same-SKU comparison across relevant local/global sources, risk notes, best link.
“I need a Zigbee USB coordinator; cheapest is fine, I can wait.”	“Where is delivery? Home Assistant/Zigbee2MQTT/ZHA, or should I compare broadly?”	Use compatibility docs plus AliExpress/global sources and local fast alternatives; compare price, chip, firmware, delivery.
“Compare robot vacuums under $400 with fast delivery to Berlin.”	“Any must-have features: self-emptying, mopping, pet hair, or mapping?”	Region-aware EU/local sources, clean queries, market map, priced recommendations.
“Посмотри Samsung 990 Pro 2TB”	“Куда доставка? Аналоги тоже проверить, если найдутся лучше по отзывам/цене?”	Verify Samsung first; analogs only if allowed.
“Сандалии кожаные детские для девочки, 26, срочно”	“Куда доставка? Бюджет? Закрытый нос/примерка важны?”	Clean query “сандалии детские”; filters: 26, delivery, budget, material, closed toe; use region-relevant stores.

Notes for future versions

Implement later, not required for MVP 1.0.0:

• category playbooks;
• structured JSON output schema;
• price watch and coupons;
• deterministic extraction scripts;
• image/photo search, reserved for a major 2.0 release.