game-scout（游戏侦察员）

Game Tips — Multi-Source Research Pipeline

Deliver real, tested, actionable, current gaming insight by orchestrating parallel research across multiple sources. The goal is to replace hours of searching and watching content with a single, well-sourced answer that gives the player a competitive edge.

Available Scripts

Script Options Quick Reference

exa-search.mjs: -n 10 (results count), --domain reddit.com (limit to domain), --exclude bad.com, --after 2026-01-01 (date filter), --contents (include page text), --summary (include AI summary), --category tweet|news

exa-contents.mjs: Multiple URLs supported. --summary "question" for targeted summary.

exa-similar.mjs: -n 10, --domain, --after, --contents

bright-scrape.mjs: Multiple URLs supported. --country us for geo-targeting.

bright-twitter.mjs: Pass one or more tweet URLs. --timeout 60 (wait time in seconds). Collects by URL — find tweet URLs first via Exa search.

yt-transcript.py: --no-meta to skip metadata.

Phase 1 — Query Analysis

Before searching, analyze the user's question to determine the research strategy.

Identify These Elements

1. Game: Exact title. Resolve abbreviations (PoE = Path of Exile, ER = Elden Ring, Val = Valorant, D2 = Destiny 2 or Diablo 2 depending on context, LoL = League of Legends, WoW = World of Warcraft, MH = Monster Hunter, FFXIV = Final Fantasy XIV).

1. Topic Type:

- Build/Loadout: Weapon, armor, skill, talent, or gear combinations - Strategy/Guide: How to approach encounters, modes, or progression - Mechanic/Interaction: How a specific system, item, or ability works - Meta/Tier List: What's currently strongest or most popular - Patch/Balance: Recent changes and their impact - Pro Play/Esports: What competitive or high-level players are using

1. Recency Requirements:

- Critical (live-service games with frequent patches): Must find current-patch info - Moderate (games with periodic updates): Recent info preferred, older is OK - Low (stable/single-player games): Evergreen guides are fine

1. Scope:

- Narrow ("does X proc bleed?"): Target wiki/database, skip broad search - Broad ("best builds for class X"): Full pipeline

Decision Table

Phase 2 — Parallel Discovery

Cast a wide net. Run multiple search commands to discover the best sources.

Read references/search-strategies.md for game-specific query templates and community hub URLs.

A. Exa AI Semantic Search (2-3 queries)

Exa understands intent — phrase queries naturally. Run these in parallel:

# Primary search
node {baseDir}/scripts/exa-search.mjs "best [topic] for [game] [current patch/season]" -n 10 --after 2026-01-01

# Reddit-focused
node {baseDir}/scripts/exa-search.mjs "[game] [topic] discussion recommendations" -n 5 --domain reddit.com --after 2025-06-01

# YouTube video discovery
node {baseDir}/scripts/exa-search.mjs "[game] [topic] guide tutorial" -n 5 --domain youtube.com --after 2025-06-01

B. Twitter/X (recent community takes)

First find tweet URLs via Exa, then extract full data via Bright Data:

# Step 1: Find relevant tweets via Exa search
node {baseDir}/scripts/exa-search.mjs "[game] [topic] meta" --domain twitter.com -n 5

# Step 2: Extract full tweet data for the URLs found
node {baseDir}/scripts/bright-twitter.mjs "https://x.com/user/status/123" "https://x.com/user/status/456"

Evaluate Discovery Results

From all results, identify the best 5-8 sources to extract in depth:

• Prefer recent content (check dates)
• Prefer high-engagement Reddit threads
• Prefer YouTube videos from known guide creators
• Prefer wiki/database pages for factual/stat questions
• Include at least 2 different source types for cross-referencing

Phase 3 — Deep Extraction

Go deep on the best sources. Run extraction commands for each source type.

Read references/source-extraction.md for detailed extraction patterns.

Reddit Threads

Use Bright Data scraper — it bypasses Reddit's bot detection. Prepend old. for cleaner scrapes:

node {baseDir}/scripts/bright-scrape.mjs "https://old.reddit.com/r/[sub]/comments/[id]/[slug]" "https://old.reddit.com/r/[sub]/comments/[id2]/[slug2]"

Focus on: OP content, top-voted comments, comments with specific data. Discard: jokes, deleted comments, tangents.

YouTube Videos

python3 {baseDir}/scripts/yt-transcript.py "https://www.youtube.com/watch?v=VIDEO_ID"

In the transcript, look for: section markers ("first/second/third"), build specifications, stat numbers, specific item/weapon names, caveats.

Fallback if no subtitles: scrape the YouTube page for description + comments:

node {baseDir}/scripts/bright-scrape.mjs "https://www.youtube.com/watch?v=VIDEO_ID"

Articles & Wiki Pages

Use Exa for clean extraction (works best on articles/wikis):

node {baseDir}/scripts/exa-contents.mjs "https://fextralife.com/..." "https://maxroll.gg/..." --summary "What build is recommended?"

For sites that block Exa, fall back to Bright Data:

node {baseDir}/scripts/bright-scrape.mjs "https://example.com/guide"

Context Management

Summarize extracted content as you go. For each source, distill to:

• Key recommendations/findings
• Specific data points (stats, percentages, item names)
• Date/patch version
• Source URL for attribution

Phase 4 — Synthesis & Validation

Cross-reference extracted information to deliver validated insights.

Cross-Reference Protocol

1. Identify consensus: Do 3+ sources agree? That's high confidence.
2. Spot conflicts: If sources disagree, note both perspectives and explain why (different patch versions, skill levels, game modes).
3. Check recency: Is the advice from the current patch/season? If a source predates a relevant patch, flag it.
4. Validate specifics: If a build claims specific stats, verify against wiki/database data when possible.

Confidence Assessment

Red Flags

• Source is from a previous patch and the topic is affected by balance changes
• Reddit thread has top comments disagreeing with OP
• YouTube video has comments saying "this was nerfed"
• Conflicting info between wiki and community — community is usually more current

Phase 5 — Structured Response

Deliver the answer in a format adapted to the query type.

Universal Structure

1. TL;DR: One to three sentences with the direct answer
2. The Details: Actionable specifics (build specs, step-by-step, tier placements)
3. Why This Works: The underlying mechanic or synergy that makes it effective
4. Caveats: Patch dependency, skill floor, mode-specific, rank-dependent considerations
5. Sources: Numbered list with links, creator names, and dates
6. Confidence: HIGH/MEDIUM/LOW with brief reasoning

Format by Query Type

Build/Loadout queries: Use a table or structured list with alternatives.

Meta/Tier list queries: Use S/A/B tier format with explanations per tier.

Mechanic/Interaction queries: Direct yes/no answer first, then detailed breakdown.

Strategy/Guide queries: Numbered step-by-step with reasoning per step.

See examples/sample-queries.md for full format examples of each type.

Fallbacks & Error Handling

If a script fails, degrade gracefully rather than abandoning the research.

If a search returns no relevant results, broaden the query or try alternative phrasing before giving up.

Reference Files

• references/search-strategies.md — Read when formulating search queries. Contains game-specific community hubs, subreddit names, and query templates by topic type.
• references/source-extraction.md — Read when extracting content from sources. Contains patterns for Reddit parsing, YouTube transcript processing, and wiki extraction.
• references/game-databases.md — Read when the query involves specific game data or interactive build planners. Contains URLs and navigation hints per game.
• examples/sample-queries.md — Read for calibration. Shows the full pipeline applied to four different query types.