research-engine

research-engine

Deep research via Claude Code slash commands or the Codex skill. Give it a URL or a topic — it returns a structured markdown report with citations.

Installation

From the published marketplace (recommended):

claude plugin marketplace add gprecious/gprecious-marketplace   # one-time per machine
claude plugin install research-engine@gprecious-marketplace

Then restart Claude Code or run /plugins reload. Updates via claude plugin update research-engine@gprecious-marketplace once a new version is published.

Codex plugin install

This repository also contains a Codex-compatible plugin manifest and skill:

.codex-plugin/plugin.json
skills/research-engine/SKILL.md

Install it the same way you install local Codex plugins in your environment, pointing the plugin source at this repository. After the plugin is available, ask Codex for research directly, for example:

Use research-engine to research https://arxiv.org/abs/2402.10171 --yes
Use research-engine to research "MoE LLM routing improvements" --yes
Use research-engine to follow up on the latest session: "저자 한계만 더 정리해줘"

Local development install (for contributors)

Clone this repo, then register it as a local marketplace:

git clone https://github.com/gprecious/research-engine
# Point a local marketplace at the repo
mkdir -p research-engine-mp/.claude-plugin
cat > research-engine-mp/.claude-plugin/marketplace.json <<JSON
{
  "\$schema": "https://anthropic.com/claude-code/marketplace.schema.json",
  "name": "research-engine-local", "version": "0.0.0",
  "owner": { "name": "local-dev" },
  "plugins": [{ "name": "research-engine", "source": "../research-engine", "category": "productivity" }]
}
JSON
claude plugin marketplace add ./research-engine-mp
claude plugin install research-engine@research-engine-local

Tip: once installed, replace the snapshot in ~/.claude/plugins/cache/research-engine-local/research-engine/<ver>/ with a symlink to your working copy so local edits reflect without a version bump.

Usage

Claude Code slash commands

/research https://youtu.be/xxxxx          # analyze a YouTube video
/research https://arxiv.org/abs/2301.xxxx # analyze a paper
/research "MoE LLM trends in 2026"        # topic research
/research https://... --yes               # skip intent Q&A
/research https://... --fresh             # bypass cache

/research-followup "저자 배경 더"         # ask follow-up on latest session
/research-followup "..." --slug <name>    # target specific session

/research-visualize <slug>                                 # charts + Notion auto-push
/research-visualize <slug> --diagrams                      # + Mermaid diagrams (preset-themed when --preset set)
/research-visualize <slug> --slides                        # + Marp slide deck (.pptx/.pdf)
/research-visualize <slug> --slides --judge                # + 4-axis rubric auto-refine <75 (2-pass cap)
/research-visualize <slug> --slides --preset dark-neon     # force preset (chart + deck + diagram share it)
/research-visualize <slug> --slides --brand-image <url>    # watermark chart background via QuickChart plugin
/research-visualize <slug> --no-sync-notion                # skip Notion push
/research-visualize                                         # use most recent session

/research-design <slug>                                    # research → claude.ai/design → LXC 배포
/research-design <slug> --no-deploy                        # 로컬 산출물까지만
/research-design <slug> --login-headful                    # cloak skip, Tailscale m4 로 1회 로그인
/research-design <slug> --fresh                            # storageState 캐시 무시

Output lands in research/YYYY-MM-DD-<slug>/README.md. When Notion is configured, a consolidated report is also upserted as a row in a research-engine database under the configured parent page (one row per session, re-runs update in place).

AI agent instructions

Agents that cannot execute Claude Code slash commands should use skills/research-engine/SKILL.md as the entrypoint and treat the command files as the canonical pipeline reference:

• commands/research.md defines the full classify → preview → intent → evidence → synthesize → persist workflow.
• commands/research-followup.md defines session follow-ups.
• lib/adapter_contract.md defines the normalized evidence JSON shape.
• lib/report_sections.md defines the required markdown report sections and citation rules.
• agents/*-adapter.md defines source-specific collection behavior.

For a new research run, create research/YYYY-MM-DD-<slug>/README.md, sources.json, and intent.json. Use the same report structure as lib/report_sections.md; every factual claim in 핵심 포인트, 상세 분석, and 인용 / 원문 must cite a real source id such as [1].

For Codex specifically: