deep-research

name	deep-research
description	Multi-source deep research with structured analysis and mandatory QA verification. Produces sourced, entity-by-entity style reports with executive summaries, taxonomies, and detailed breakdowns. Use when user says deep research, investiga [topic], research [topic] for [client], analisis de mercado, competitive analysis, benchmark [topic], or needs a comprehensive market/product/regulatory investigation. Do NOT use for quick factual lookups (use WebSearch directly) or for content creation (use seo-content or social-content instead).
metadata	{"author":"Alfonso Sainz de Baranda (Growth4U)","version":"1.0","system":"Growth Raistlin"}

Deep Research

Investigación multi-fuente estructurada en 7 fases. De pregunta vaga a informe verificado con fuentes.

When to Use

• Market analysis for a client (banks, competitors, regulations)
• Product benchmarking (pricing, features, models)
• Regulatory / legal landscape research
• Technology landscape evaluation
• Any research that will be shared with a client or stakeholder

Workflow Overview

SCOPE → SOURCES → EXTRACT → FRAMEWORK → DETAIL → QA → DELIVER
  1        2         3          4          5       6      7

Each phase produces an artifact. Never skip a phase. The output of each feeds the next.

---

Phase 1: SCOPE

Define exactly what we're researching before touching any source.

Inputs to Capture

Field	Example (Overdrafts)
Research question	How do Spanish banks handle overdrafts? Opt-in vs tacit?
For whom	Mauricio (Monzo Spain GTM)
Entities to cover	All major retail banks in Spain (ING, BBVA, CaixaBank, N26...)
Data points per entity	Model type, activation, limits, costs, grace period, sources
"Complete" means	Every bank with >1M customers in Spain covered
Output format	Markdown report with executive summary + bank-by-bank breakdown
Client folder	01-business/clients/[client]/research/

Rules

• If user hasn't specified scope clearly, ask with AskUserQuestion before proceeding.
• Write scope as a 3-5 line brief at the top of the working document.
• Scope defines the stopping criteria — without it, research never ends.

---

Phase 2: SOURCE DISCOVERY

Find ALL relevant sources before extracting any data.

Search Strategy

Run minimum 5 different search queries varying:

• Language (Spanish + English for Spain topics)
• Angle (product name, regulatory term, comparison term, consumer forum)
• Source type (official sites, comparison platforms, regulators, news, forums)

Source Categories

Priority	Type	Example
1	Official	Bank websites, regulator (BdE, CNMV), company docs
2	Comparison	Rankia, Roams, HelpMyCash, Kelisto, BusconOmico
3	News	El Economista, CincoDías, Expansión, Bloomberg
4	Legal/Regulatory	BOE, Tribunal Supremo rulings, EU directives
5	Community	Reddit, Forocoches, Twitter/X, specialized forums

Output

Source inventory list: URL + what it likely contains + reliability rating (A/B/C).

Minimum 10 sources before proceeding. If <10, run more search queries.

Phase 2b: SOCIAL PULSE (Conditional — /last30days)

Activate when the research topic has social/community dimension, recency matters, or user sentiment is relevant.

Activate	Skip
Community perception matters	Regulatory/legal research
Topic is trending or emergent	Historical analysis
Need market buzz beyond official data	Product feature comparison from official docs
User sentiment drives the insight	Topic too niche for social media

Process:

1. Run python3 ~/.claude/skills/last30days/scripts/last30days.py "[research topic]" via Bash (foreground, 5-min timeout)
2. The script searches Reddit, X (Twitter), YouTube, and web for the last 30 days
3. Integrate findings as a "Social Pulse" source category (Priority 5b) in the source inventory
4. Feed social sentiment data into Phase 3 extraction alongside official sources

If unsure, ask the user: "This topic may benefit from a social pulse check (Reddit, X, YouTube last 30 days). Should I run it?"

---

Phase 3: DATA EXTRACTION

Extract structured data from each source.

Process

1. Open each source (WebFetch or WebSearch for key claims)
2. Extract data points matching the scope definition
3. Use consistent field structure across all entities
4. Note source URL for EVERY data point
5. Mark confidence: verified (official source) / reported (secondary) / inferred (deduced)

Rules

• One claim, one source minimum. No unsourced claims.
• When sources conflict, note BOTH versions and flag for Phase 6 (QA).
• Extract raw data first — don't synthesize yet.
• For numerical data: always note the date/year of the data.

---

Phase 4: FRAMEWORK & TAXONOMY

Create the organizing structure that makes the data intelligible.

Process

1. Review all extracted data
2. Identify natural groupings, patterns, categories
3. Create a taxonomy or framework that explains the landscape
4. Build comparison tables

Example (from Overdrafts research)

The data revealed 3 models, not the assumed 2:

1. Opt-In (customer activates) — ING, N26, Openbank
2. Opt-Out (bank activates, customer can disable) — BBVA
3. Pure Tacit (no control) — Santander, Sabadell, etc.

Output

• Taxonomy diagram or table
• Summary table (entities as rows, key dimensions as columns)
• Key insight: what's the non-obvious finding?

Rule

The framework should surprise the reader with a non-obvious insight. If it just confirms what everyone already knows, dig deeper.

---

Phase 5: DETAILED ANALYSIS

Entity-by-entity deep breakdown using the framework.

Structure per Entity

Use a consistent template across all entities:

### [Entity Name] — [Product Name]

**Type:** [Category from taxonomy]

**How it Works:**
- [Step-by-step user journey]

**Key Data:**
| Feature | Detail |
|---------|--------|
| [Dimension 1] | [Value] |
| [Dimension 2] | [Value] |

**Sources:**
- [Source 1](URL)
- [Source 2](URL)

Document Structure

# [Research Title]

**Date:** YYYY-MM-DD
**For:** [Stakeholder name and role]
**Research by:** Alfonso Sainz de Baranda (Growth4U)

---

## Executive Summary
[Key findings table + 2-3 paragraph narrative]

---

## [Taxonomy/Framework Section]
[Models, categories, comparison tables]

---

## Detailed Analysis

### Category A: [Name]
#### 1. [Entity] — [Product]
#### 2. [Entity] — [Product]

### Category B: [Name]
...

---

## Recommendations
[Actionable next steps for the stakeholder]

## Sources Index
[All sources cited, grouped by type]

Rules

• Every entity gets the SAME template structure (no entity gets less coverage than another).
• Executive summary must be standalone — readable without the detail sections.
• Include a "Recommendations" section with actionable takeaways for the stakeholder.

---

Phase 6: QA VERIFICATION (Mandatory)

This phase is NOT optional. Every deep research MUST be QA'd.

Process

1. Save the analysis document to the client research folder
2. Invoke /qa-bot on the saved document
3. The QA bot will:
   • Extract all factual claims
   • Generate 10-15 verification questions (Deep QA mode)
   • Independently verify each claim via web search
   • Compare against the document
   • Produce a QA Report with confidence score

After QA

QA Verdict	Action
PASS (score >=9/10)	Proceed to Phase 7
NEEDS REVISION (7-8.9/10)	Fix flagged issues, re-run QA on fixes only
MAJOR ISSUES (<7/10)	Rework affected sections, full QA re-run

Rules

• Fix ALL errors and discrepancies before delivering.
• For UNVERIFIABLE claims: either find a source, mark explicitly as "no public data available", or remove the claim.
• Save the QA report alongside the analysis: QA-REPORT-[filename].md

---

Phase 7: DELIVER

File Outputs

Save to 01-business/clients/[client]/research/:

File	Content
[topic]-analysis.md	The full detailed analysis
QA-REPORT-[topic]-analysis.md	The QA verification report

Notion (if requested)

Create a Notion page under the client's workspace with the executive summary + link to full analysis.

Handoff

Present to the user:

1. Executive summary (inline in chat)
2. Key non-obvious finding
3. QA confidence score
4. File paths where everything is saved
5. Notion link (if created)

---

Quality Standards

Standard	Minimum
Sources per entity	>= 2 (at least 1 official)
Total unique sources	>= 10
QA confidence score	>= 8/10 before delivery
Claims without source	0 (all must be sourced or flagged)
Conflicting data	Noted with both versions
Data freshness	Year noted for all metrics