Search Triage

Iterative Search and Triage

Purpose

Perform human-like exploratory search: multiple search rounds with keyword variants, opening promising items, checking comment activity, backing out to list, continuing until budget exhausted or sufficient candidates found.

Scope

What This Skill Does

• Execute multi-round searches with keyword variants
• Open and inspect individual posts/articles
• Assess engagement signals (likes, comments, shares)
• Return to list and continue browsing
• Track visited items to avoid re-processing
• Produce ranked candidate list with evidence

What This Skill Does NOT Do

• Deep-read all comments (use comment-reading-and-sentiment-scan)
• Extract full article content (use full-article-extraction)
• Generate or post comments (use comment skills)
• Make final decisions (presents candidates for user/next skill)

Inputs

From Task Specification

Runtime Parameters

Outputs

Candidate List Format

search_results:
  task_id: string
  platform: string
  search_duration_seconds: int

  statistics:
    searches_performed: int
    items_scanned: int
    items_opened: int
    candidates_found: int

  keywords_used:
    - keyword: string
      results_count: int
      useful_results: int

  candidates:
    - rank: int
      title: string
      author: string
      url: string  # If available
      platform_id: string  # Post ID

      relevance:
        score: float  # 0-1
        keyword_matches: list
        reason: string

      engagement:
        likes: int
        comments: int
        shares: int
        engagement_rate: string  # "high" / "medium" / "low"

      content_preview: string  # First 200 chars
      has_active_discussion: boolean

      evidence:
        screenshot: string  # Path if captured
        excerpt: string

  rejected:
    count: int
    reasons:
      - reason: string
        count: int

  errors: list
  recommendations: string

Primary Workflow

Phase 1: Prepare Search Variants

Before searching, expand keywords:

Original: "AI agents"

Variants:
- "AI agents" (exact)
- "AI agent" (singular)
- "LLM agents"
- "autonomous agents"
- "Claude agent" (if relevant)
- 中文: "AI智能体" (if targeting Chinese content)

Phase 2: Execute First Search

1. Launch platform app
   → launch_and_wait(package="com.instagram.barcelona", wait_text="Search")

2. Navigate to search
   → find_and_click(text="Search") OR platform-specific location

3. Enter search keyword
   → type_and_submit(text="AI agents", submit=true)

4. Wait for results
   → smart_wait(text="post" OR text="result", timeout=5)

5. Record initial view
   → get_screen_summary()

Phase 3: Scan Results

For each visible item:
  1. Extract visible info (title, author, preview, engagement indicators)
  2. Quick relevance check:
     - Does title/preview contain keywords?
     - Is engagement above threshold?
     - Is it recent enough?
  3. Decision:
     - OPEN if: keyword match + decent engagement
     - SKIP if: no match or very low engagement
     - MARK if: maybe relevant, check later if time permits

Phase 4: Open and Inspect

For items marked OPEN:
  1. Click to open
     → find_and_click(text=item_title)

  2. Wait for load
     → smart_wait(text="comment" OR text="like", timeout=3)

  3. Scan content:
     - Read full visible text
     - Note comment count
     - Check if comments seem active (visible comment preview)

  4. Quick scroll (1-2 times) to see more content
     → mobile_swipe_on_screen(direction="up")

  5. Record findings

  6. BACK to results
     → navigate_back(expected_text="Search")

  7. Mark as VISITED to avoid re-checking

Phase 5: Scroll and Continue

After processing visible items:
  1. Scroll down to load more
     → mobile_swipe_on_screen(direction="up", distance=400)

  2. Wait for new content
     → smart_wait(timeout=1.5)

  3. Check for new items (compare with visited list)

  4. Repeat Phase 3-4 for new items

  5. Continue until:
     - max_scrolls reached
     - max_items processed
     - no new content loading

Phase 6: Next Keyword Round

If more keywords to try AND budget remaining:
  1. Clear search OR navigate to fresh search
  2. Enter next keyword variant
  3. Repeat Phases 3-5
  4. De-duplicate (same posts may appear for different keywords)

Phase 7: Rank and Report

1. Sort candidates by relevance score
2. Filter below threshold
3. Generate evidence (screenshots of top candidates)
4. Compile final candidate list
5. Add recommendations for next steps

Heuristics

Relevance Scoring

def calculate_relevance(item, keywords):
    score = 0.0

    # Title match: +0.4
    if any(kw.lower() in item.title.lower() for kw in keywords):
        score += 0.4

    # Content match: +0.3
    if any(kw.lower() in item.content.lower() for kw in keywords):
        score += 0.3

    # Engagement bonus: +0.2
    if item.engagement > 50:
        score += 0.2
    elif item.engagement > 20:
        score += 0.1

    # Recency bonus: +0.1
    if item.is_recent:  # Within time_range
        score += 0.1

    return min(score, 1.0)

When to Skip vs Open

Scroll Budget Allocation

If 3 keywords and max_scrolls=5:
  - Keyword 1: 3 scrolls (primary term)
  - Keyword 2: 1 scroll (variant)
  - Keyword 3: 1 scroll (variant)

De-duplication

Track by:

1. Post ID (if available)
2. Author + first 50 chars of content
3. Exact title match

Stop Conditions

• max_items reached
• max_scrolls reached
• max_time_minutes exceeded
• Same content appearing (infinite scroll loop)
• 3 consecutive scrolls with no new relevant items

Failure Modes & Recovery

1. Search Returns No Results

Symptom: Platform shows "No results" or empty state.

Recovery:

• Try next keyword variant
• Remove special characters from keyword
• Try broader term
• Report to user if all variants fail

2. Items Don't Open

Symptom: Clicking on item doesn't navigate to detail view.

Recovery:

• Try clicking on different part of item (title vs. image)
• Check if item is an ad (skip)
• Wait longer for tap to register
• After 2 failures, skip and note error

3. Can't Return to List

Symptom: Back press doesn't return to search results.

Recovery:

• Try pressing BACK again
• Check if in comments section (need extra back)
• If stuck, go HOME and restart search
• Note which items were lost

4. Infinite Scroll Loop

Symptom: Same items keep appearing after scroll.

Recovery:

• Check screen hash (if unchanged, stop scrolling)
• Note: "End of results reached"
• Move to next keyword

5. Login/Auth Wall

Symptom: Platform asks to log in to see more.

Recovery:

• Report: "Additional results require login"
• Stop current search
• Return found candidates so far

6. Rate Limited

Symptom: "Slow down" message or search disabled.

Recovery:

• Wait 30-60 seconds
• Reduce scroll frequency
• If persists, stop and report

Tooling (MCP)

Primary Tools (mobile-macro)

Secondary Tools (mobile-mcp)

Usage Pattern

1. Launch: launch_and_wait(package, wait_text="Search")
2. Search: find_and_click(text="Search")
3. Type: type_and_submit(text=keyword)
4. Get elements: mobile_list_elements_on_screen
5. Click item: find_and_click(text=item_title)
6. Read: get_screen_summary()
7. Back: navigate_back()
8. Scroll: mobile_swipe_on_screen(direction="up")
9. Repeat 4-8

Examples

Example 1: Basic Threads Search

Input:

platform: threads
keywords: ["AI agents", "LLM agents"]
max_items: 10
max_scrolls: 5

Execution Log:

[1] Launched com.instagram.barcelona
[2] Found search icon (top-right area)
[3] Searched "AI agents"
[4] Results loaded - 6 visible posts
[5] Scanned posts:
    - Post 1: "My AI agent failed..." - 23 likes, relevant → OPEN
    - Post 2: "Best travel agents in NYC" - 45 likes, not relevant → SKIP
    - Post 3: "Building agents with Claude" - 89 likes, relevant → OPEN
[6] Opened Post 1, read content, 12 comments, BACK
[7] Opened Post 3, read content, active discussion, BACK
[8] Scrolled, 4 new posts visible
[9] ...continued...
[15] Switched to "LLM agents" keyword
[16] 3 new results (2 duplicates filtered)
[20] Budget exhausted, compiling results

Output:

search_results:
  platform: threads
  search_duration_seconds: 420

  statistics:
    searches_performed: 2
    items_scanned: 18
    items_opened: 7
    candidates_found: 5

  candidates:
    - rank: 1
      title: "Building agents with Claude"
      author: "@techdev"
      relevance:
        score: 0.9
        keyword_matches: ["AI agents", "Claude"]
        reason: "Direct keyword match, high engagement, active discussion"
      engagement:
        likes: 89
        comments: 34
        engagement_rate: "high"
      has_active_discussion: true

    - rank: 2
      title: "My AI agent failed - lessons learned"
      author: "@aibuilder"
      relevance:
        score: 0.7
        keyword_matches: ["AI agent"]
        reason: "Experience report, moderate engagement"
      engagement:
        likes: 23
        comments: 12

    # ... more candidates ...

  recommendations: "Candidate #1 has most active discussion, recommend for comment-reading skill"

Example 2: WeChat Article Search

Input:

platform: wechat
keywords: ["AI大模型", "人工智能"]
scope: articles
max_items: 8

Note: WeChat search works differently - use "搜一搜" feature.

Output:

search_results:
  platform: wechat

  candidates:
    - rank: 1
      title: "2024年AI大模型发展趋势报告"
      author: "机器之心"
      relevance:
        score: 0.95
        reason: "Exact keyword match, authoritative source"
      content_preview: "本报告详细分析了2024年上半年AI大模型领域的..."

Example 3: No Results Scenario

Input:

platform: threads
keywords: ["xyznonexistentterm123"]

Output:

search_results:
  platform: threads

  statistics:
    searches_performed: 1
    items_scanned: 0
    candidates_found: 0

  keywords_used:
    - keyword: "xyznonexistentterm123"
      results_count: 0
      useful_results: 0

  candidates: []

  errors:
    - type: NO_RESULTS
      keyword: "xyznonexistentterm123"
      message: "No results found for this keyword"

  recommendations: "Try different keywords. Suggestions: broaden the search term or check spelling."

Key Reminders

1. Human-like browsing - Don't just scan metadata; actually open and look at items
2. Track visited items - Never process same item twice
3. Keyword variants matter - First keyword may miss relevant content
4. Back to list - Always return to results after viewing item
5. Evidence for candidates - Each candidate needs reason + engagement data
6. Respect budgets - Stop when max_items/scrolls/time reached
7. De-duplicate across keywords - Same post can appear in multiple searches
8. Report rejected items - Helps calibrate thresholds