deepdive
view on github →
§ claude code skill · v2.0 · MIT licensed

Stop ad-hoc
Googling.
Start documented
investigation.

A 9-phase meta-research skill for Claude Code. Hypothesis testing, parallel sub-agent search, source triangulation, adversarial review. Output is a folder you can return to a month later — every claim traces to a source file with quotes and scoring.

install in 30s → github
9
Workflow phases
29
Search channels
105
Report blocks
460+
Stat sources
39+
API endpoints
§ 01 · methodology

Why this exists

Most AI research is one prompt, one wall of text, no way to verify or reuse. Deepdive turns that into a process you can audit.

REJECTED

Without this

  • One-shot prompt, no plan, no documented decisions
  • Sources lost in chat history — no return path
  • No way to detect confirmation bias
  • "Sources include..." — vague, unverifiable
  • Generic Google results, no methodology diversity
  • No reuse next research — start from zero
VERIFIED

With this

  • 17-section plan.md documents every choice
  • Each source = file with verbatim quotes + scoring
  • Mandatory adversarial pass + opposition queries
  • Every claim → [s12] link → specific quote
  • 29 named channels with paywall fallback protocols
  • Atomic theses in findings/FN.md — reusable
// session-log · claude-code · deepdive
investigate trade-offs between Postgres logical replication and CDC tooling
Reframed: 3 falsifiable hypotheses formulated
Picked genre: decision (comparison + validation)
Wrote plan.md (17 sections, time-boxed: 2h)
Checked env: 4 APIs available, 2 fallback to HTML
Launched 4 parallel Explore sub-agents
Saved 23 sources to sources/ with quotes + scoring
Ran adversarial pass (3 steel-man counter-arguments)
REPORT READY → research/postgres-replication-vs-cdc/2026-05-21_decision.md
§ 02 · workflow

Nine phases.
Every one transparent.

Each phase has a defined output and a checkpoint. You confirm key decisions. The skill records what it chose and why in plan.md.

1

Reframing

Restates the question. Identifies the decision it supports. Formulates 2-4 falsifiable hypotheses.

2

Genre

Picks report type: qa, explainer, decision, landscape, validation, or custom.

3

Plan

Writes plan.md with 17 sections: acceptance criteria, risk register, sourcing strategy.

3.5

Capabilities

Audits available API keys. Maps subtopics to data sources. Surfaces gaps before execution.

4

Search

Launches 2-5 parallel Explore sub-agents across 29 channels and 460+ stat sources.

5

Score

Each source rated Credibility/Recency/Bias. Every claim backed by ≥3 independent sources of different types.

6

Synthesis

Assembles the report from blocks, then runs an adversarial pass: 4 opposition questions, steel-man counter-arguments documented.

6.5

Verify

Lightweight citation check before closing. Confirms every claim still resolves to a saved source quote.

7

Refresh

Extracts entities, numbers, and hypotheses into refresh_targets.md — the entry point for a later delta-update without re-running everything.

§ 03 · catalog

A curated catalog.
Auto-validated weekly.

460+ statistical sources, 39+ API endpoints, 29 named channels, 105 report blocks. Weekly cron in GitHub Actions validates endpoints and discovers upstream additions.

/blocks
105 / 10 categories

Report Blocks

Reusable sections with templates, anti-patterns, and composition rules. Each block has a fixed shape — you compose a report by naming the blocks it contains.

tldr mental-model data-table counter-arguments +71 more
/channels
29 strategies

Search Channels

Named search strategies with query patterns and paywall fallback protocols. Each channel documents what it covers and where it breaks.

academic preprint-servers code-github crypto-analytics +25 more
/stat-sources
460+ sources · 33 categories

Stat Sources

Curated catalog with URL · Access · Quality · Limitations · Combine-with · Fallback. 14 cross-industry plus 19 specific industries.

FRED SEC EDGAR WHO CoinGecko +276 more
/genres
6 + custom

Report Types

Genre defines structure. Five standard presets cover most cases. Custom assembles per question from the block library.

qa explainer decision landscape validation custom
/api-sources
39+ endpoints

API Catalog

Free no-auth APIs prioritized. Auth via env vars only — skill never asks for keys inline. Documented fallback strategies.

Semantic Scholar OpenAlex DefiLlama PubMed +26 more
/automation
weekly cron

Auto-Validation

GitHub Actions validates endpoints and discovers upstream awesome-list additions. Auto-PR for dead endpoints. Reports in a dedicated branch.

HEAD checks upstream sync auto-PR
§ 04 · install

Install in 30 seconds.

Works on Claude Code (CLI), Claude Desktop with Skills enabled, and any other LLM with manual context loading.

// recommended

Claude Code (CLI)
git clone https://github.com/Socialpranker/\
  deepdive.git \
  ~/.claude/skills/deepdive
  • Type "deep research" or "deep dive"
  • Works across all projects
  • Auto-loads via progressive disclosure

// .skill bundle

Claude Desktop
git clone ...
cd deepdive
zip -r ../deepdive.skill . \
  -x ".*" -x "*.zip"
  • Upload via Settings → Skills → Add
  • Appears in Customize panel
  • Triggers same as CLI

// manual load

Other LLMs
# Load SKILL.md + references/*.md
# into context manually
# ~70% LLM-agnostic markdown
  • Codex, Gemini, local models
  • Sub-agents → separate chat sessions
  • PRs welcome for adapters
§ 05 · questions

Frequently asked.

How is this different from ChatGPT Deep Research or Perplexity?
Those are products — closed UI, fixed flow, opaque source selection. This is open methodology — you control every step, the protocol is markdown you can fork, the source catalog is yours to extend. ChatGPT and Perplexity also don't separate sources into files, don't do explicit triangulation, don't run adversarial passes, and don't produce reusable atomic theses.
Does it work without Claude Code CLI?
Yes — on Claude Desktop with Skills enabled. Also works manually with any LLM by loading the markdown files into context. Sub-agent parallelism needs to be simulated with separate chat sessions per subtopic.
Why so many files? Isn't this overkill?
For a 5-minute "what's the latest X" question — yes. That's why shallow mode exists (5-7 sources, no sub-agents, ~15 min). The full machinery is for medium (1 hour) and deep (3 hours) when you need to use the output for a decision. The file-per-source structure is the reuse mechanism — a single research often informs 3-5 future researches.
What if I don't have CLAUDE.md or a project context?
The skill detects context in 3 tiers: explicit (CLAUDE.md research_root setting) → autodetect (pyproject.toml, package.json) → fallback (~/deep-research/). No project, no problem.
Is this just prompt engineering?
It's structured methodology plus a curated catalog plus reusable templates plus automation. The 9-phase workflow forces discipline. 460+ stat sources is curated knowledge. 105 reusable blocks compose any report shape. Weekly auto-validation keeps the catalog alive. 25+ upstream awesome-lists give a discovery layer. Prompts are an implementation detail, not the value.
Can I use this commercially?
Yes — MIT licensed. Use it, modify it, integrate it into products. Attribution appreciated but not required.

Star it.
Use it.
Extend it.

The catalog grows through contributions. Easiest path: add a stat source you know to the right industry file. 15-minute PRs welcome.

star on github how to contribute