Indagatio

1. WebFetch-first for all web URLs: ALWAYS try WebFetch before Playwright. Most sites (docs, blogs, news) are SSR and WebFetch gets full content. Playwright ONLY after WebFetch returns empty/broken content requiring JS rendering. See references/tool-selection.md Playwright Gate.
2. Multi-pass extraction: WebFetch over-summarizes; always use multiple passes
3. Verification mandatory: Never trust constructed URLs; validate before output
4. Scale-appropriate depth: Small/Medium/Large determines thoroughness
5. Bilingual output: Korean with English terms in parentheses
6. Tool-specific handling: Each source type has optimal tool selection
7. Cross-source verification: Every source needs complementary sources checked
8. Scope expansion: Proactively discover related topics user didn't think of
9. Quality goal: Produce HIGHER quality than manual curation (if output is narrower → scope expansion insufficient)
10. Adaptive batching: Dynamically batch large results to prevent hallucination

Self-check: Is this more comprehensive than manual? Did I discover unknown info? If not → redo Step 2b.

Cross-source recommendations:

Source Types & Tools

→ See references/source-detection.md for URL patterns. → See references/tool-selection.md for tool decision tree and prompts. → See ../_common/references/frontmatter-standard.md for status taxonomy.

Operation Mode

Memory Check → Source Detect → Scale → Fetch (multi-pass + link crawl) → Verify → Analyze → Quick Note → Log → Memory Save

indagatio ends after Quick Note creation. Daily Note logging handled by purgatio.

Core Workflow

Step 0: Memory Check

Check claude-mem for cached topic info before fetching externally.

mcp__plugin_claude-mem_mcp-search__search: query: "{topic}", limit: 5

If found → ask user: "캐시된 정보가 있습니다 (검증일: YYYY-MM-DD). 새로 분석할까요?" If not found → proceed to Step 1.

Step 1: Source Detection

→ See references/source-detection.md for regex patterns and detection order.

1a. Multi-source: Same topic → merge; different topics → separate notes; independent → parallel.

1b. Detection order: file extension → URL patterns → special prefixes → Obsidian link → vault path → general URL → topic search.

1c. Duplicate Check: Search 00. INBOX/01. Quick Notes/Indagatio-Log.md for URL. If found → ask user.

1d. MCP & Dependency Check:

• skill-seekers not connected → notify user with install command, offer WebFetch fallback
• multi-model-ai not available for YouTube → notify user, skip video analysis

1e. Existing Note Check:

1. Search vault: Glob: **/*{topic}*.md + Grep: {topic} in vault path
2. If found → present options: A) New note (with comparison after), B) Enhance existing, C) Cancel
3. If A → provide comparison after completion (sources count, scope, unique info, merge recommendation)

Step 2: Scale Assessment

→ See references/scale-algorithm.md for decision tree.

User Override: --small, --medium, --large, --no-crawl (skip link crawl), --no-expand (skip scope expansion).

Step 2b: Scope Expansion (MANDATORY)

After assessing primary source, ALWAYS search for related topics:

1. Brave Search (1순위, Pro Plan: 50/sec parallel, unlimited):

   • General: "{topic} guide tutorial" count: 15
   • Community: "{topic} site:reddit.com OR site:news.ycombinator.com" count: 10
   • FAQ: "{topic} FAQ troubleshooting" count: 10
   • News (if relevant): "{topic} 2025 2026 update release" count: 10
   • 429 error → WebSearch fallback

2. Present scope expansion suggestion to user with discovered subtopics

3. If confirmed → fetch additional sources per subtopic

Reddit note: Use old.reddit.com (www blocks Playwright).

Skip: --no-expand, highly specific single API endpoint, explicit scope limit.

Step 2c: Bulk Result Processing

When Topic Search / Scope Expansion returns 10+ results → hierarchical processing to prevent hallucination.

Skip when: <10 items, single source, --no-batch, total tokens < 30K.

→ See references/hierarchical-summarization.md for batch calculation, hierarchy, and processing rules.

Step 3: Content Fetching (MULTI-PASS — MANDATORY ALL PASSES)

Run ALL applicable passes in sequence. Never skip a pass. Pass 1b is NOT optional for Medium/Large.

3a. Pass 1 — Structure + Links (all scales): WebFetch with structure+links extraction prompt. Extract headings, word counts, content types, and ALL inline hyperlinks as a LINK INVENTORY.

3b. Pass 1b — Related Link Crawl (MANDATORY for Medium/Large): Never skip when scale ≥ Medium. This step is required.

1. Parse link inventory from Pass 1 output
2. Filter: exclude anchors (#), nav, auth, assets → score (same-domain+shared path=3, same-domain=2, external=1) → select top N (Medium: max 3, Large: max 6)
3. Fetch each selected link with single-pass WebFetch
4. Apply Atomicity Gate: distinct concept (1500+ words, independent meaning, different scope) → separate quick note + [[wikilink]]; supporting detail → inline
5. Set links_crawled in frontmatter to actual count fetched

This step takes ~34s per link (~100-200s total). No excuse to skip. Skip ONLY when: Small scale, --no-crawl flag, non-web source, or 0 links in inventory.

→ See references/tool-selection.md Pass 1b for full procedure, filtering rules, and Atomicity Gate criteria.

3c. Pass 2 — Detail (Medium/Large): WebFetch with detail extraction prompt. Incorporate context from Pass 1b linked pages.

3d. Pass 3 — Images (if images detected in Pass 1): WebFetch with image URL extraction prompt. Verify CDN patterns.

Tool selection order for web URLs:

1. WebFetch (DEFAULT) — use for ALL web URLs including docs, blogs, news, articles
2. Playwright (FALLBACK ONLY) — ONLY when WebFetch returns empty/broken content. See references/tool-selection.md Playwright Gate.
3. Doc frameworks (Mintlify, Docusaurus, GitBook, Nextra, VitePress, MkDocs) are SSR → WebFetch always works

Key rules:

• WebFetch does NOT return actual image URLs → use explicit extraction prompt (Pass 3d)
• If image URLs not extractable → placeholder ![Figure N: Description](URL_NOT_AVAILABLE)
• YouTube: Prompt user to run Gemini CLI /indagatio <URL> manually → provide output file → parse
• NEVER call gemini CLI from Claude Code Bash (Google ToS violation)
• Reddit: always old.reddit.com (exception: Playwright allowed for Reddit)

→ See references/tool-selection.md for all prompts, code examples, and fallback workflows.

Step 4: Verification (MANDATORY)

4a. Image URL Verification: Check URL pattern validity (CDN patterns OK, hex strings you generated → INVALID). Invalid → placeholder + warning.

4b. Content Completeness:

If thin → re-fetch with more specific prompts.

4c. Language Balance: Technical terms need English in parentheses. Concepts explained, not just listed.

4d. Comment Insights (Reddit, HN, forums only):

Filter comments by quality (technical detail, specific data, corrections to OP → High; practical tips → Medium; opinions without evidence → Low).

Cross-verify every extracted comment via Brave Search/official docs:

• ✅ Verified (2+ sources), ⚠️ Partial (1 source), ❌ Unverified, 🚫 Debunked

→ See references/quicknote-format.md for 댓글 인사이트 output section format.

Skip: Non-community sources, <5 comments, no High quality comments, --no-comments.

4e. Difficulty Assessment:

Assess content difficulty (1-10) based on source_type × scale baseline:

Content complexity adjustment: +1 if multi-domain, +1 if requires prior knowledge, -1 if introductory. Clamp to 1-10.

Add difficulty: {value} to quick note frontmatter after scale:.

Step 5: Content Analysis & Writing

Bilingual Term Rule: 한국어 용어 (English Term) — all technical terms, proper nouns, abbreviations, tool names.

Depth: Small → bullets OK; Medium → 2-3 sentences/concept; Large → full paragraph with examples.

AI Processing: Structure → Key concepts (full explanations) → Why it matters → Practical examples → Action items → Vault connections.

Step 6: Quick Note Generation

CRITICAL: Timestamp Acquisition — NEVER guess. Always run before creating file:

date "+%Y.%m.%d-%Hh%Mm"  # Filename
date "+%Y-%m-%d %H:%M"   # Frontmatter

Output Path: 00. INBOX/01. Quick Notes/notes/YYYY.MM.DD-HHhMMm {Source Title} Analysis.md

→ See references/quicknote-format.md for full template.

Step 7: Log Recording

Append to 00. INBOX/01. Quick Notes/Indagatio-Log.md:

| YYYY-MM-DD | HH:mm | <source_url> | <source_type> | <scale> | <image_status> | [[Quick Note Title]] |

Step 8: Memory Storage

Save verified information to claude-mem after Quick Note creation:

mcp__plugin_claude-mem_mcp-search__save_memory:
  title: "{topic}", text: "핵심 정보...\n출처: {URL}\n검증일: YYYY-MM-DD"

Skip: Vault folder review, image-only analysis, failed fetches.

Step 9: Completion Verification

Before reporting completion, verify:

1. Quick note file created at expected path (file must exist)
2. links_crawled > 0 in frontmatter (Medium/Large web source)
3. 관련 페이지 section present with populated table (Medium/Large)
4. Atomicity Gate decisions made for each crawled link
5. Log entry appended to Indagatio-Log.md
6. Memory saved to claude-mem (if applicable)

If any verification fails → fix before reporting completion.

User Confirmation Points

1. Duplicate URL found
2. Scale override needed
3. MCP fallback
4. Output location (project-related → suggest project folder)
5. Image extraction failed

Quality Checklist

• Images: CDN URLs (not hallucinated)
• Content depth meets scale minimum
• Korean + English terms
• Concepts explained, not listed
• Cross-source: complementary sources suggested
• Comment insights cross-verified (community sources)
• Scope expanded: discovered related topics
• Related links: key cross-references followed (Medium/Large)
• Better than manual curation

If unchecked → back to Step 2b.

Vault Folder Review

Special workflow for reviewing internal vault folders (projects, system guides, etc.).

→ See references/vault-folder-review.md for full template and analysis guidelines.

Key differences from external: Data from internal vault; focus on status + progress (not content extraction); output includes expansion suggestions based on external resources.

Expansion suggestions: Context-dependent analysis of current level, project purpose, tech stack, trends → recommend next-level resources, related tools, practical applications, communities.

Integration

• Inbound: User direct call via /indagatio or natural language
• Outbound: Quick Notes processed by purgatio; URLs logged in Indagatio-Log.md

Reference Files

• references/source-detection.md - URL pattern matching
• references/scale-algorithm.md - Scale determination
• references/quicknote-format.md - Output template
• references/tool-selection.md - Tool decision tree and prompts
• references/vault-folder-review.md - Vault folder review template
• references/hierarchical-summarization.md - Bulk result processing
• references/reddit-guide.md - Reddit navigation
• ../_common/references/logging-standard.md - Unified logging

Output Format Rules

SKILL.md content: English (context efficiency) User-facing output: Korean (user preference) Trigger patterns: English + Korean (accessibility)