resumator

Multi-source document analysis platform generating structured reports with configurable detail levels. Supports 7 content types (meeting, study, watch, training, comparison, technical, audit) and 5

SKILL.md

GitHub

Resumator v3.0 — Multi-Source Analysis Platform

Overview

Resumator transforms raw content (transcriptions, articles, documents, URLs) into structured, exhaustive reports. It supports 7 content types and 5 detail levels, with optional web enrichment and strict source traceability.

Core principle: Zero hallucination — every claim is sourced or marked uncertain.

Quick Start

1. User provides source(s)
2. Resumator asks: Type? Level?
3. User responds (e.g., "b5" for Study/Exhaustive)
4. Resumator processes and generates report

Workflow

Step 1: Receive Sources

Accept any combination of:

  • Meeting transcripts (with speaker names)
  • YouTube transcripts (with timecodes)
  • URLs (articles, documentation)
  • PDFs / uploaded documents
  • Raw notes

Step 2: Configuration Questionnaire (ALWAYS ASK)

After receiving sources, ALWAYS present:

📊 Resumator v3 — Configuration

1️⃣ Type de traitement ?
   a. 📋 Réunion — Compte-rendu structuré
   b. 🔬 Étude — Recherche approfondie  
   c. 📰 Veille — Synthèse d'actualités
   d. 📖 Formation — Extraction pédagogique
   e. ⚖️ Comparatif — Analyse comparative
   f. 🔧 Technique — Documentation tech
   g. 📊 Audit — Analyse critique

2️⃣ Niveau de détail ? (1-5)
   1. ⚡ Flash — TL;DR en 5 lignes
   2. 📋 Résumé — Points clés (~500-800 mots)
   3. 📊 Détaillé — Analyse complète (~1500-2500 mots)
   4. 📚 Approfondi — + contexte (~3000-5000 mots)
   5. 🔬 Exhaustif — Recherche maximale (5000+ mots)

💡 Raccourci : tape "a3" pour Réunion/Détaillé ou "b5" pour Étude/Exhaustive

→ See questionnaire.md for shortcuts and defaults

Step 3: Process Based on Type

Type Web Search Template
📋 Réunion Only if level ≥4 reunion.md
🔬 Étude Yes (exhaustive) etude.md
📰 Veille Yes (news focus) veille.md
📖 Formation If level ≥4 formation.md
⚖️ Comparatif Yes (alternatives) comparatif.md
🔧 Technique Yes (docs) technique.md
📊 Audit If level ≥4 audit.md

→ See niveaux-detail.md for level specifications

Step 4: For Types Requiring Web Search

Execute research workflow:

  1. Extract themes and gaps from sources
  2. Generate search plan (5-7 axes)
  3. Execute searches iteratively
  4. Evaluate source reliability (1-5 ⭐)
  5. Fetch and extract relevant content
  6. Cross-check contradictions

→ See workflow-recherche.md for full process

Step 5: Integrate Multiple Sources

When multiple sources provided:

  1. Normalize each source to common structure
  2. Build thematic cross-index
  3. Detect contradictions
  4. Merge with full traceability

→ See integration-sources.md for fusion rules

Step 6: Generate Report

  1. Use appropriate template for selected type
  2. Apply detail level constraints
  3. Include source citations [N] and [🌐N]
  4. Add confidence scores where relevant
  5. Output as downloadable .md file

File naming: [TYPE]_[YYYY-MM-DD]_[slug].md

  • CR_2025-01-13_reunion-gardel.md
  • ETUDE_2025-01-13_claude-code.md

Critical Rules

🔴 CRITICAL (Never Violate)

  1. QUESTIONNAIRE: ALWAYS ask type + level before processing
  2. ANTI-HALLUCINATION: Every claim has source [N]/[🌐N] or marked "⚠️ non vérifié"
  3. EXHAUSTIVITY: Every action/decision/commitment mentioned MUST be captured
  4. FIDELITY: Never invent. If uncertain → "à confirmer" or "non spécifié"
  5. TRACEABILITY: Web enrichments marked with 🌐, skill additions with ⚠️
  6. LANGUAGE: Output in same language as input

🟡 IMPORTANT

  1. TL;DR FIRST: Even level 5, executive summary at top (mandatory)
  2. SOURCE PRIORITY: Primary > Recent > Secondary
  3. CONTRADICTIONS: Document both sides, don't arbitrarily choose
  4. PROGRESS FEEDBACK: For long processing, show progress updates
  5. DIAGRAMS: Auto-detect flows → Mermaid (max 6 per report)

🟢 DESIRABLE

  1. GLOSSARY: Extract acronyms and technical terms
  2. RELATED TOPICS: Suggest areas for further exploration
  3. KEY QUOTES: Preserve impactful verbatims

→ See anti-hallucination.md for detailed rules


Source Traceability (Perplexity-style)

Inline Citations

Claude Code allows terminal-based task delegation [1]. 
It supports slash commands for navigation [2][🌐1].

Bibliography Section

## 📖 Sources
### Fournies
[1] Transcript YouTube "Claude Code Tutorial" — 03:45
[2] Documentation PDF — page 12

### Recherches web
[🌐1] docs.anthropic.com — "Slash commands reference"
[🌐2] blog.anthropic.com — "Claude Code announcement"

### Non retenues
- oldsite.com — Obsolète (2022)

Reliability Scores

Score Meaning
⭐⭐⭐⭐⭐ Official docs, primary sources
⭐⭐⭐⭐ Recognized tech media, verified experts
⭐⭐⭐ Community (SO, Reddit with arguments)
⭐⭐ Unsourced blogs, undated
⭐ Obsolete, unreliable → excluded

Meeting Mode (📋 Réunion) — Preserved from v2

For backward compatibility, meeting reports retain v2 structure:

Meeting Type Detection

Type Indicators
Steering/Decision "décision", "valider", budget, deadlines
Information "informer", "présenter", updates
Brainstorming "idées", "explorer", creative
Technical "architecture", "API", technical terms
Generic No clear indicators → neutral plan

Action Status Indicators

Indicator Meaning
🟢 Owner AND deadline defined
🟡 Owner OR deadline (one missing)
🔴 Neither owner nor deadline

→ See templates/reunion.md for full template → See legacy/mermaid-detection.md for diagrams → See legacy/proactive-rules.md for insights → See legacy/glossary-extraction.md for glossary


User Options

Option Effect
--no-web Disable web search
--no-diagrams Disable Mermaid generation
--no-glossary Disable glossary extraction
--max-diagrams N Limit diagram count (default: 6)

Note: Options can be stated naturally ("sans recherche web", "pas de diagrammes")


Knowledge Base

Configuration

Workflows

Templates (7 types)

Legacy (from v2)


Limitations

This skill does NOT:

  • Perform audio transcription (speech-to-text)
  • Translate content
  • Generate meeting agendas
  • Process video content directly
  • Generate more than 6 diagrams per report
  • Access paywalled content

Version History

Version Date Changes
1.0.0 2025-12-10 Initial release
2.0.0 2025-12-16 Added: Mermaid diagrams, proactive insights, glossary, action indicators
3.0.0 2025-01-13 Added: 7 content types, 5 detail levels, questionnaire, web research, multi-source integration, anti-hallucination, Perplexity-style citations

Current: v3.0.0

Owner

  • Author: Édouard
  • Contact: Via Claude.ai