Socialpranker/deepdive

Deepdive – a Claude Code skill for structured meta‑research

What it is – Deepdive is a skill for Anthropic’s Claude Code that turns a plain “research X” request into a reproducible, 13‑phase pipeline. The pipeline automatically:

1. Re‑frames the question and generates hypotheses.
2. Chooses a report genre (e.g., decision, explainer) and a set of search channels.
3. Builds a detailed plan and lets the user approve it before any search runs.
4. Dispatches parallel sub‑agents (using Claude Haiku) to fetch sources from 29 curated channels (web, academic, code, regulatory, etc.).
5. Triangulates claims, filters evidence, and runs a CRAG‑style relevance classifier.
6. Synthesizes a report, adds an adversarial “red‑team” pass, and verifies citation faithfulness (RAGAS/ALCE‑style).
7. Walks the user through decision forks and logs the outcome for future reuse.

All artefacts are saved as markdown files (plan, sources, claims, numbers, memo, etc.) so the research can be revisited months later and the provenance of every claim is explicit.

---

Quick install (30 s)

For Claude Code – clone the repo into ~/.claude/skills/:

git clone https://github.com/Socialpranker/deepdive.git ~/.claude/skills/deepdive

Then type commands like “Investigate post‑gres logical replication vs CDC tooling” inside a Claude Code session.

For Claude Desktop – clone, zip the folder as a .skill bundle, and upload via Claude.app → Settings → Skills.

Other LLMs – the 13‑phase methodology is portable; load SKILL.md and the reference markdown files into the model’s context and run the phases manually.

---

How it works (high‑level)

Phase	Name	Core task
1	Reframing	Classify question type and generate hypotheses
2	Genre & block selection	Pick report genre and block templates
3	Plan	Produce a 17‑section execution plan
3.5	Capability discovery	Detect which model (Opus/Haiku/Sonnet) to use
3.7	Plan‑review gate	Show plan to user for approval/edit
4	Search	Parallel sub‑agent queries across 29 channels
5	Claims ledger & triangulation	Record each claim, check source diversity
5.5	Evidence filter	CRAG‑style relevance filter per claim
5.7	Wiki reconcile	Merge overlapping evidence
6	Synthesis + red‑team	Write report, generate adversarial counter‑arguments
6.5	Verify	Faithfulness check (link alive, entailment, qualifiers)
7	Refresh targets	Add missing gaps, re‑run limited searches
8	Decision walkthrough	Interactive fork resolution, log outcome

Model routing is explicit: high‑reasoning phases use Claude Opus, cheap fan‑out uses Haiku, synthesis uses Sonnet. Cost is shown up‑front and can be forced into “cheap mode” if desired.

---

Key assets shipped with the repo

• Report blocks – 106 pre‑written block templates (FRAME, EXPLAIN, COMPARE, etc.)
• Search channels – 29 named strategies with query patterns and pay‑wall fallbacks
• Statistical sources – 460+ curated APIs and datasets (gov, finance, health, crypto, etc.)
• API catalog – 47 free, no‑auth endpoints (Semantic Scholar, arXiv, World Bank, SEC EDGAR, …)
• Eval harness – scripts to benchmark runs on six quality dimensions (citation integrity, cost, accuracy, coverage, adversarial honesty, determinism)
• Validation scripts – check_citations.py, check_number_provenance.py, scripts/validate_phases.py enforce that every phase produced its expected artefacts.

---

What you get after a run

research/<topic>/
├─ plan.md                # 17‑section execution plan
├─ claims.csv             # ledger of every claim + triangulation status
├─ sources/…/*.md         # one file per source with verbatim quotes
├─ findings/F*.md         # atomic theses with confidence levels
├─ .verify/*.json         # liveness, faithfulness, qualifier checks
├─ memo.md                # one‑page decision memo (recommendation, risks)
└─ report.md              # full multi‑section research report

Each claim is linked to its source ([s12]) and to verification results, making the output fully auditable and reusable.

---

Who might use it

• Product managers or engineers who need a documented, bias‑checked investigation before making a technical decision.
• Researchers who want a reproducible evidence‑gathering pipeline without manually tracking citations.
• Teams that already use Claude Code and want to extend it with a “research‑as‑code” workflow.

---

Where to learn more

• Docs site – https://socialpranker.github.io/deepdive/
• Quickstart – QUICKSTART.md
• Evaluation harness – eval/README.md
• Contribution guide – CONTRIBUTING.md

---

Bottom line

Deepdive is a fully‑featured, open‑source Claude Code skill that automates rigorous, citation‑rich research. By breaking a query into 13 transparent phases, it reduces hallucinations, surfaces bias, and produces a reusable, version‑controlled research folder that can be revisited or audited later.