Socialpranker/deepdive

What it solves

Deepdive is a structured meta-research skill for Claude Code that replaces ad-hoc searching and one-shot prompting with a rigorous, documented investigation pipeline. It prevents common AI research failures such as hallucinated citations, source bias, and the "wall of text" output that lacks a clear decision-making path.

How it works

The project implements a 13-phase pipeline that guides an LLM through a systematic research process:

1. Reframing & Planning: It transforms a user's query into a decision specification with hypotheses and a detailed 17-section plan. It uses a router to choose the best decomposition method based on the question's shape (e.g., comparative vs. factual).
2. Targeted Search: It employs parallel sub-agents across 29 named search channels (academic, legal, code, etc.) and utilizes a registry of over 460 statistical sources and 47+ APIs. It includes a "no-progress circuit breaker" to stop wasting tokens on dead ends.
3. Triangulation & Scoring: Claims are logged in a ledger and scored for credibility, recency, and bias. It specifically flags "input-level skepticism" (e.g., vendor-reported benchmarks) before they reach synthesis.
4. Verification: The system runs multiple machine-checked passes, including a CRAG-style relevance filter and a faithfulness check to ensure the source actually entails the claim.
5. Synthesis & Execution: It produces a decision memo and a full report. The process concludes with a mandatory "decision walkthrough" where the user executes the findings against their original decision forks.

Who it’s for

Developers and researchers using Claude Code or other LLMs who need high-confidence, verifiable research reports that lead to concrete actions rather than generic summaries.

Highlights

• 13-Phase Pipeline: A rigorous workflow from reframing to final decision walkthrough.
• Multi-Model Routing: Optimizes cost and quality by routing tasks to Opus, Sonnet, or Haiku based on the complexity of the phase.
• Automated Verification: Includes scripts to check citation liveness, claim entailment, and the arithmetic of derived numbers.
• Extensive Source Registry: Built-in access to dozens of specialized APIs and hundreds of of industry-specific statistical sources.
• Decision-Centric Output: Focuses on producing a "decision spine" (memo and forks) rather than just a summary of information.