deep-research

Six composable research skills (L0→L5). From a 60-second fact-check to a 40-minute academic-grade investigation with 120+ sources. Pick the right depth for the task.

What is this?

A set of six Claude Code skills that form a research ladder. Each tier composes on top of the previous one, adding specific capabilities:

Specialized skills (parallel to ladder)

Key property: L3 calls L2 which calls L1. Every higher level inherits the foundation and adds its own layer (reflection, critic, multi-perspective, academic sources, recursive loops).

Why does this exist?

LLM-assisted research suffers from three chronic problems:

• Shallow by default — ad-hoc "search the web" gives you 1–2 sources and confident-sounding synthesis with no verification.
• Hollow synthesis — the output sounds authoritative, but there is no audit trail. You can't trace a claim back to a source.
• Context bloat — pouring research into your main conversation bloats the context window and degrades later tool use.

Quick Start

Prerequisites

• Claude Code CLI (any recent version)
• Firecrawl CLI — scraping engine
• Tavily MCP — search API (free tier works)

Optional add-ons (each skill degrades gracefully if absent):

• Exa MCP — neural semantic search (L2+, v0.5.0+) · free tier 1000/mo
• Perplexity MCP — answer engine (L0+ killer path, v0.6.0+)
• OpenAI Codex CLI — cross-model channel (L2+, v0.2.0+) · needs ChatGPT Pro

Full setup: see docs/INSTALLATION.md.