deep-research

---
name: deep-research
description: Use when conducting rigorous, multi-source investigations backed by ≥3 independent sources per claim, parallel sub-agents, adversarial review, and persistent source documentation.
category: Research
version: 1.0.0
tools: []
---

# Deep Research — Disciplined Meta-Research

Turn research into an auditable, reusable investigation: every claim traces to a specific source file, documented assumptions, and a refresh protocol for future updates.

## Phases Overview

| # | Phase | Purpose |
|---|-------|---------|
| 1 | Reframe | State 2–4 falsifiable hypotheses |
| 2 | Genre & blocks | Pick report type and structure |
| 3 | Plan | Write scope, sourcing strategy, risk register |
| 4 | Search | Parallel sub-agents, fetch & dedup, save sources |
| 5 | Score & triangulate | Rate credibility, require ≥3 independent sources per thesis |
| 6 | Synthesize + adversarial | Assemble report, run self-critique, add counter-arguments |
| 7 | Refresh targets | Extract entities/numbers for delta-updates |

## Core Mechanisms

**Triangulation:** Every thesis backed by ≥3 independent sources of different types (primary/academic/industry/discussion). Claims with fewer flagged "insufficient evidence."

**Source-grounding:** Each source becomes `sources/NN_slug.md` with metadata, verbatim quotes, scores. No dangling claim—every assertion links to a specific file.

**Adversarial pass:** Phase 6 runs strongest reasoning: 4 self-critique questions plus active search for counter-arguments and disconfirming evidence.

**Falsifiable hypotheses:** Phase 1 commits to 2–4 hypotheses; Phases 5–6 explicitly confirm/refute each against evidence or mark under-determined.

**Parallel sub-agents:** Phase 4 launches search sub-agents concurrently (cheap models for broad web, stronger for reasoning-heavy subtopics).

**Refresh protocol:** Phase 7 emits `refresh_targets.md`; an `update <slug>` run produces delta (new entrants, entity changes, refreshed numbers, adversarial triggers) instead of full replay.

## Output Structure

```
<slug>/
├── plan.md                    # scope, sourcing strategy, risk register
├── sources.csv               # index with scores
├── sources/
│   ├── 01_<slug>.md         # one source per file with metadata + verbatim quotes
│   └── ...
├── findings/
│   └── F1_<short>.md        # reusable theses
├── refresh_targets.md        # what to watch on update
├── diffs/
│   └── YYYY-MM-DD_delta.md  # delta from update run
└── YYYY-MM-DD_<genre>.md    # final report
```

## When to Use

- Low-quality answer is expensive: strategy, business plan, report, article groundwork.
- Comparing N institutions/products/methods/markets requiring defensible reasoning.
- Validating hypothesis or decision against external data.
- Meta-research: understanding how X works, mapping landscape of Y.

## Anti-Patterns

Do not skip reframing or existing-work check. Do not output to chat only—persist sources and report to files. Do not fabricate citations; if fetch returns nothing, claim is empty. Do not build on thin corpus; say so rather than overstating. Do not skip adversarial pass. Do not run sub-agents sequentially—fan-out in parallel. Do not collapse sources into one file; per-source files enable reuse.