Agentic Recipes

Practical, end-to-end playbooks for working on UnoPim with AI. Each recipe states the goal, which agentic interface it uses, the steps (the prompts and tools involved), and what to verify afterwards.

The three interfaces these recipes draw on:

• AI Agent — the in-app chat widget (reference).
• MCP Server — your IDE's AI assistant connected to the live instance (reference).
• Agentic Skills — UnoPim conventions loaded into your coding agent (reference).

---

Recipe 1 — Scaffold a Connector with AI

Goal: Stand up a new third-party connector plugin without hand-writing boilerplate.

Uses: Agentic Skills (unopim-plugin-development) + MCP Server (dev_tools → generate_plugin, run_command).

Steps:

1. With the unopim-plugin-development skill installed, prompt your coding agent:

   "Scaffold a connector plugin named 'ShipStation' with a configuration page, credential storage, and a connection test."

2. The skill supplies UnoPim's package conventions (service provider, Database/Migration/ folder, ['admin'] middleware, repository pattern); the MCP generate_plugin action writes the directory structure, composer.json, and base classes into packages/Webkul/ShipStation.
3. Run migrations through the bridge:

   "Run php artisan migrate." → MCP run_command.

Verify: Ask the agent to read_file the generated ServiceProvider, then confirm the plugin's config page loads in the admin panel. Run vendor/bin/pint --test.

---

Recipe 2 — Bulk Attribute Enrichment

Goal: Fill missing descriptions across hundreds of products.

Uses: AI Agent (search + content generation + bulk edit + auto-translation).

Steps:

1. "Find products without a description created in the last 24 hours." → the agent calls SearchProducts.

2. "Generate a 200-word SEO-friendly description for each from its name and attributes." → GenerateContent.

3. "Apply the descriptions and translate them to all active locales." → BulkEdit, then the TranslateProductValuesJob runs on the queue.

Verify: Spot-check a few products in the admin panel; confirm locale-dependent fields populated for each active locale. If approval_mode is review, approve the changeset from the approval queue.

TIP

For products created before auto-translation was enabled, run the CLI alternative: php artisan unopim:translate.

---

Recipe 3 — Catalog Completeness Diagnostics

Goal: Find why a product isn't appearing in a channel and fix the gaps.

Uses: MCP Server (get_product, get_catalog_schema, upsert_products).

Steps:

1. "Check product SKU 'PHONE-001' completeness for channel 'default'. If it's below 100%, list the missing required attributes."

2. The assistant fetches the product with its completeness data and reports the missing required attributes.

3. "Fill the missing attributes with sensible values and upsert." → upsert_products (max 50 per atomic batch).

Verify: Re-run the completeness check; confirm 100% for the target channel. Confirm the product now surfaces in that channel.

---

Recipe 4 — Mass Category Re-assignment

Goal: Restructure the catalog — move a set of products to a new category branch.

Uses: AI Agent (SearchProducts, AssignCategories, BulkEdit).

Steps:

1. "Find all products in the 'Winter Wear' category."

2. "Assign them to 'Seasonal > Winter' and remove the old 'Winter Wear' assignment."

Verify: Confirm the count of re-assigned products matches the original search count; check the category tree in the admin panel. Use review approval mode for large moves so the changeset is auditable and reversible.

---

Recipe 5 — Define Attributes by Description

Goal: Rapidly prototype catalog structure during development.

Uses: MCP Server (upsert_attributes, upsert_attribute_groups) or AI Agent (CreateAttribute, ManageFamilies).

Steps:

1. "Create a select-type attribute 'Fabric Material' with options Cotton, Silk, Wool."

2. "Add it to the 'Clothing' attribute group."

Verify: Confirm the attribute and its options exist via search_attributes, and that it renders on a product in the relevant family.

---

Recipe 6 — AI-Assisted Test & Review Loop

Goal: Ship a change that passes the UnoPim development pipeline.

Uses: Agentic Skills (unopim-dev-cycle, unopim-code-review) + MCP Server (generate_test, run_command).

Steps:

1. "Generate a Pest test for the new ShipStationRepository." → MCP generate_test (or mcp:make test).

2. "Run the test." → run_command → vendor/bin/pest.

3. "Review my changes against UnoPim standards before I open a PR." → the unopim-code-review skill flags convention violations (table prefix, route middleware, hardcoded strings, repository pattern).

Verify: Tests pass, vendor/bin/pint --test reports zero issues, and php artisan unopim:translations:check passes with zero missing keys.

---

Recipe 7 — Reusable Workflow as a Skill

Goal: Stop re-typing a multi-step instruction you run often.

Uses: Dynamic Skills (MCP) — see Extending the MCP Bridge.

Steps:

1. Drop a SKILL.md into .ai/skills/audit-channel-pricing/ describing the steps (search products, compare prices across two channels, report mismatches).
2. The bridge registers execute_audit_channel_pricing.
3. Invoke it any time instead of re-explaining the sequence.

Verify: Run the new tool from your AI editor and confirm it produces the expected mismatch report.

---

General Best Practices

• Discover before acting — call get_catalog_schema first so the assistant knows your custom attributes and filterable fields.
• Keep batches atomic — upsert_* tools cap at 50 items per call for clean error reporting.
• Use review mode for bulk changes — queue AI-driven mutations as changesets so they're auditable and reversible before they hit the catalog.
• Sequence read → write → verify — read the current state, make the change, then confirm with a follow-up search or completeness check.
• Let skills enforce conventions — keep the relevant Agentic Skill active so generated code respects UnoPim's standards the first time.