prakhar1114/ai_mime_marketplace/find-restaurants-and-send-on-whatsapp

Find Restaurants and Send on WhatsApp

Searches Google Maps for a restaurant query, extracts the top place names and URLs (default 5), closes the Maps search tab, then sends an intro plus each place as a separate message to a WhatsApp contact using the native WhatsApp Mac app.

Inputs

• search_query (required, string): The Google Maps search, typically cuisine + area

(e.g. "italian restaurants in indiranagar").

• contact_name (required, string): Exact WhatsApp contact name to open and message

(e.g. "Prakhar Jain"). A self-chat appears as "<name> (You)" and matches fine.

• result_limit (optional, integer): Number of top unique restaurant results to extract

and send. Defaults to 5.

Run

Run via the executable bash wrapper:

./run.sh [path/to/inputs.json]

If no path is given it defaults to inputs/inputs.example.json.

Preconditions: the user must be logged into whatsapp app and AI Mime app needs to be running.

Python runtime contract:

• run.sh uses the first available interpreter in this order: skill .venv/bin/python,

workflow .venv/bin/python, then the required $AI_MIME_PYTHON_PATH.

• scripts/run.py uses only the Python standard library, so no virtualenv is required.
• The skill shells out to $AI_MIME_BROWSER_HARNESS_BIN for Google Maps and

$AI_MIME_UI_AGENT_CMD for native WhatsApp.

How It Works

1. find_top_restaurants (browser_harness): Opens

https://www.google.com/maps/search/<urlencoded query> in a new Chrome tab, waits for div[role=feed], reads unique result cards (div.Nv2PK) using .qBF1Pd for the name and a.hfpxzc[href] for the Google Maps URL, then closes the Maps tab.

1. send_whatsapp_message (ui_agent): Activates native WhatsApp, confirms the target

chat, sends suggested places to eat, do a thumbsup where we should meet, then sends each place as its own single-line message: <index>. <name> - <url>. The UI agent uses clipboard paste and the Return/enter key; it verifies the first send and final state but avoids slow per-message screenshots.

Outputs

• restaurant_places (list[dict]): Top restaurants in display order, each with name

and url.

• restaurant_messages (list[str]): The intro message followed by one message per place.
• sent_confirmation (dict): { "sent": true, "messages_sent": <count>, "note": "..." }

from the UI agent.

Progress Log Format

The script emits JSON log events on stderr:

• {"event": "step_start", "id": "find_top_restaurants", "title": "..."}
• {"event": "step_done", "id": "find_top_restaurants", "outputs": {...}, "summary": "..."}
• {"event": "step_failed", "id": "...", "error": "...", "recoverable": false}
• {"event": "workflow_done", "outputs": {...}}

The process exits non-zero on step_failed.

Fallback

If Google Maps extraction fails or WhatsApp cannot complete, follow references/fallback_plan.md. The Maps fallback is manual/UI-agent extraction from the left results panel; the WhatsApp fallback is already UI-agent driven.

ask_llm Decision Points

None. Restaurant extraction is deterministic DOM scraping, and the WhatsApp flow is handled by the UI agent. No ask_llm calls are used.

References

• fallback_plan.md: Manual/UI-agent recovery steps.
• learned_notes.md: Durable selectors, URL pattern,

WhatsApp activation notes, and speed gotchas.