learn-platform

name	learn-platform
description	Discover how to automate a new app or website. Explore UI elements, scrape docs, build selector references. Use when: "learn how to automate", "explore the UI of", "map the interface", "discover selectors", "build reference for", "what can you do in this app", "how does this app work", "explore app".
allowed-tools	["mcp__sh__platform_explore","mcp__sh__platform_learn","mcp__sh__platform_guide","mcp__sh__apps","mcp__sh__windows","mcp__sh__ui_tree","mcp__sh__browser_dom","mcp__sh__browser_page_info","mcp__sh__browser_navigate","mcp__sh__screenshot_file","mcp__sh__browser_open","mcp__sh__memory_recall","mcp__sh__memory_save","mcp__sh__memory_query_patterns"]

Platform Discovery & Learning

You are exploring a new app or website to build automation knowledge — discovering UI elements, selectors, keyboard shortcuts, and flows.

Intelligence Wrapper

Every tool call returns automatic hints. When you navigate to a URL or focus an app, the server auto-loads any matching reference file and injects selector hints, known errors, and strategy suggestions into responses. Watch for [HINT], [WARNING], and [STRATEGY] lines.

Check Existing Knowledge First

Always start with:

platform_guide(platform="{name}")
memory_recall(task="explore {name}")
memory_query_patterns(scope="{name}")

Existing reference files (use these exact names with platform_guide):

• Social: x-twitter, instagram, linkedin, reddit, threads, youtube, discord
• Design: figma, canva-smoke-test
• Video: davinci-resolve-keyboard, davinci-resolve-menu-map, davinci-resolve-menus-batch1 through batch4
• Dev: codex-desktop, n8n, devto, devpost
• Google: google-flow

If a reference already exists, review it before exploring further. You may not need to re-explore.

Two Discovery Modes

Mode 1: Interactive Exploration (platform_explore)

Discovers all interactive elements by actively testing them.

For web apps:

platform_explore(url="https://app.example.com", maxElements=30)

What it does:

1. Navigates to the URL
2. Finds all interactive elements (buttons, links, inputs, menus, tabs)
3. Filters out dangerous elements (delete, logout, etc.)
4. Tests each element (click, check what changed)
5. Compiles a reference JSON with working selectors, navigation map, errors

For native apps:

platform_explore(bundleId="com.example.App")

What it does:

1. Launches the app if not running
2. Maps the full AX tree
3. Records all interactive elements (buttons, menus, text fields)
4. Does NOT auto-click native elements (safety)
5. Compiles reference with element roles, titles, paths

Output: Saved to references/{platform}.json.

Mode 2: Documentation Scraping (platform_learn)

Learns from official documentation without interacting with the app.

platform_learn(platform="{name}", url="https://docs.example.com/shortcuts")

What it does:

1. Generates likely doc/help URLs from the platform name
2. Navigates Chrome to each URL
3. Extracts text content, keyboard shortcuts (table parsing), CSS selectors
4. Compiles into a reference with shortcuts, features, API endpoints, limitations

Best doc URLs to provide:

• Keyboard shortcuts page
• Getting started / quick start guide
• API documentation
• Accessibility / developer docs

Output: Saved to references/{platform}.json.

Manual Exploration Workflow

When the automated tools aren't enough, explore manually:

For native apps:

1. apps → find the target app
2. focus(bundleId) → bring to front
3. ui_tree(pid, maxDepth=3) → get overview of top-level elements
4. ui_tree(pid, maxDepth=8) → deep dive into specific areas
5. screenshot_file → visual reference
6. Test interactions: ui_press, menu_click, key
7. Record what works

For web apps:

1. browser_navigate(url) → open the app
2. browser_dom(selector="button, a, input, [role='button'], [role='tab']") → all interactive elements
3. browser_page_info → page structure
4. Test key selectors with browser_click
5. Navigate to different pages, repeat
6. screenshot_file at each major section

Combining Both Modes

For comprehensive coverage:

1. platform_learn first → get keyboard shortcuts and documented features
2. platform_explore second → discover actual UI elements and test them
3. The results merge into a single reference file

Output Quality

A good reference file contains:

• platform: Name for matching
• bundleId or urlPatterns: For automatic context loading
• selectors: Grouped by feature area (navigation, compose, settings, etc.)
• flows: Named step sequences for common tasks
• errors: Known issues with context and solutions
• shortcuts: Keyboard shortcuts with descriptions

Save

After exploration:

• The reference is auto-saved to references/{platform}.json
• Call memory_save(task="explored {platform}") to record the exploration strategy
• The reference is immediately usable by platform_guide and the context tracker

$ARGUMENTS