Prompt recipes
Copy-pasteable prompts that build commerce-AI features fast — paste each into Claude (or any MCP host) with the discoverGPT MCP server connected, or adapt it to the REST API. Every recipe is grounded in real MCP tools and /v1 endpoints.
These are ready-to-paste prompts for building commerce-AI features on discoverGPT. Each recipe pairs a short goal with a prompt you can drop straight into Claude (or any MCP-aware host) that has the discoverGPT MCP server connected, plus a note on exactly which MCP tools and REST endpoints it exercises. Every prompt references only tools and endpoints that actually exist — nothing invented.
Two ways to run every recipe
With MCP — connect the discoverGPT MCP server once (browser consent, no token to paste) and paste the prompt into your host. Setup is in MCP transport; the tool list is in Tools exposed over MCP.
With REST — each recipe lists the /v1 endpoints its tools map to, so you can
hand the same steps to your own backend. Auth for direct REST is the
client-credentials grant.
Either way, every new account is auto-authorized for the read-only demo merchant
98334572911, so the prompts below return live results with no catalog upload.
How to use these
Connect the MCP server
One-command setup for Claude Code, Cursor, VS Code, Codex, and generic hosts — authorize once, tools appear.
See a working example
The live /demo trimodal search widget runs real search_products_trimodal queries against the demo catalog.
Adapt to REST
Prefer your own code? Each recipe maps to /v1 endpoints you can call with a Bearer token.
Keep tokens server-side
The REST paths below are called from your backend with a short-lived Bearer
token. Never ship a client_secret or a JWT to the browser — proxy searches and
composes through a server route, exactly as the storefront wiring
example shows.
1. Build a trimodal search box
Stand up a storefront search feature that ranks by words, a precise color,
and an image — the same capability the live /demo widget uses.
You are helping me build a product-search feature for a storefront. Use the
discoverGPT MCP tools against the demo merchant (98334572911):
1. Call search_products_trimodal for "linen shirt for a summer wedding" and show
the top 5 results as a table of name, brand, price, and product_id.
2. Re-run the same query but add a precise color of deep sage green
(HSV roughly [130, 40, 55]) and describe how the ranking shifts.
3. Then write me a minimal React <SearchBox> component that posts the query to my
own backend /api/search route (which proxies POST /v1/search) and renders the
returned discovery[] array. Keep the Bearer token server-side.Exercises search_products_trimodal (MCP) → POST /v1/search.
Working reference implementation: the live /demo widget. Note that a
non-null query_image_url dominates and ignores text/color
in the same call — run separate calls to combine modes.
2. Ship a shopping assistant that builds complete outfits
A brand-aware assistant that finds a hero piece, verifies it, and rounds out a full look with complementary products.
Act as a personal shopping assistant for our store. First ground yourself: call
get_merchant_brand_context so you match the brand's voice and know the catalog.
Then, against the demo merchant (98334572911):
1. From the occasion I give you, use search_products_trimodal to find 3 candidate
hero pieces under my budget.
2. For the one I pick, call get_product_details to confirm material, fit, and
availability before you recommend it.
3. Call recommend_complementary_products on that product_id and assemble a full
outfit (top / bottom / shoes / accessory), explaining each pairing in one
sentence, in the brand's voice.
My occasion: "outdoor autumn wedding, smart-casual, budget $350."Exercises get_merchant_brand_context, search_products_trimodal,
get_product_details, recommend_complementary_products (MCP). This flow is
MCP-native — brand context and recommendations are agent tools; the search step
maps to POST /v1/search. It mirrors
Demo 1 — NL shopping agent.
3. Add virtual try-on to a product page
Wire the one usage-metered capability — compose a shopper image wearing a garment — into a PDP, and understand what's free vs. billed before you ship.
Help me add virtual try-on to a product detail page. Using the discoverGPT MCP
tools on the demo merchant (98334572911):
1. Call get_vto_credits and tell me how many free / prepaid composes I have left
before this starts costing money.
2. Call virtual_try_on to dress subject "model-aria" in dress "SKU-1001"
(source: "pdp") and give me the returned image URL.
3. Explain the caching rule: which repeat calls are free vs. metered (~$0.18),
and how to force a fresh render.
4. Sketch the PDP button flow: on "Try it on", POST /v1/vto/compose from my server
with the shopper's subject id and this product in the dress slot, then swap the
product image for the returned path.Exercises virtual_try_on, get_vto_credits (MCP) →
POST /v1/vto/compose, GET /v1/vto/credits. Cache hits
and subject-only lookups are free; only a live (cache-miss) compose is metered.
refresh: true forces a fresh, billed render.
4. Run an AI-visibility (GEO) audit
Turn AI-discovery visibility into a prioritized action list — where your products do and don't surface inside ChatGPT, Gemini, and Perplexity.
Run an AI-visibility (GEO) audit of our catalog and give me a prioritized action
list. Using the discoverGPT MCP tools on the demo merchant (98334572911):
1. Call get_visibility_score for the catalog-wide score and the per-platform
breakdown (ChatGPT / Gemini / Perplexity).
2. Call get_visibility_trends over the last 30 days — is visibility rising or
falling, and on which platform?
3. Call get_visibility_products to list the products that are NOT visible in AI
assistants.
4. Call get_geo_recommendations and turn the gaps into a ranked to-do list:
product, the platform it's missing from, and the specific fix.
Format the result as a short summary paragraph followed by a table.Exercises get_visibility_score, get_visibility_trends,
get_visibility_products, get_geo_recommendations (MCP) →
GET /v1/visibility/score, /v1/visibility/trends,
/v1/visibility/products, /v1/visibility/geo-recommendations.
5. Check catalog quality before you publish feeds
Catch thin, low-quality, or unenriched records — and confirm the AI-discovery feeds are compiled — before anything reaches a storefront or an agent.
Audit our catalog data quality before we publish feeds. Using the discoverGPT MCP
tools on the demo merchant (98334572911):
1. Call get_catalog_quality_overview and summarize how many products sit in each
quality band and the average score.
2. Pick the 5 lowest-scoring products and call get_product_quality on each — show
the before/after content and the specific issues flagged.
3. Call get_feed_status to check whether the AI-discovery feeds (ACP, JSON-LD,
Perplexity, llms.txt) are compiled and current.
4. Give me a punch list of what to fix, ordered by impact on feed readiness.Exercises get_catalog_quality_overview, get_product_quality,
get_feed_status (MCP) → GET /v1/quality/overview,
/v1/quality/product/{product_id}, GET /v1/feed/status.
6. Browse the catalog over UCP
Explore the catalog the way an agent does under the Universal Commerce Protocol — discover, search, look up, and fetch products by query instead of downloading a feed file.
Explore our catalog through the Universal Commerce Protocol (UCP) — the way an
agent discovers and queries products rather than downloading a feed. Using the
discoverGPT MCP tools on the demo merchant (98334572911):
1. Call get_ucp_profile and tell me what the discovery profile advertises
(endpoints, capabilities).
2. Call ucp_catalog_search for "waterproof hiking boots under $150" and return the
UCP Product records.
3. For the top hit, call ucp_get_product by its id and show the full UCP record.
4. Call ucp_catalog_lookup with two or three ids at once to demonstrate batch
retrieval. Then explain when I'd query UCP live vs. download a static feed.Exercises get_ucp_profile, ucp_catalog_search, ucp_get_product,
ucp_catalog_lookup (MCP) → GET /v1/ucp/profile,
POST /v1/ucp/catalog/search, /v1/ucp/catalog/product, /v1/ucp/catalog/lookup.
Next steps
Tools exposed over MCP
The full 30-tool set behind these recipes, with a one-line description of each.
Demos
Three copy-paste POCs — an NL shopping agent, image-similarity search, and a drop-in search widget.
Search API
Request/response schemas for POST /v1/search, the endpoint most of these recipes build on.