recursive-research

Claude Code plugin & skill for recursive research up to PhD level on any topic — science, tech, business, arts, humanities. Source tiering, loop auto-regulation, disk checkpointing, and WDM + Munger inversion for autonomous decisions.

Version: 2.2.0 · License: MIT · Author: Joseph Huayhualla (@Anjos2)

What it does

You give it a research seed (a topic) and the skill:

1. Asks for mode (web / local / mixed), local paths if applicable, priority/excluded sources, and a cycle cap.
2. Identifies 3-5 seed threads applying WDM (Weighted Decision Matrix) + Munger Inversion.
3. Detects available MCPs (Firecrawl, Context7, WebFetch, WebSearch) and prioritizes by speed and quality.
4. Iterates in auto-regulated cycles — each cycle picks the least-covered thread, selects sources, investigates, consolidates.
5. Tiers every source into Tier 1 / 2 / 3 / Rejected with transparent criteria.
6. Saves disk checkpoints every cycle — survives context compaction.
7. Closes when the 5-criteria PhD fitness function is met, or upon hitting the cycle cap.
8. Asks if you want to keep going. Research can be infinite.

Why it's different

Installation

This repo is a Claude Code marketplace containing one plugin (recursive-research). The easiest way to install is through Claude Code's built-in plugin manager.

Option A — Via marketplace (recommended)

Inside Claude Code, run:

/plugin marketplace add Anjos2/recursive-research
/plugin install recursive-research

The plugin appears in /plugin → Installed. Invoke it with /recursive-research:recursive-research.

Option B — Via the official Plugin Directory (pending approval)

Anthropic maintains a central Plugin Directory. Once approved there (submission in progress), install becomes a one-liner without the marketplace add step:

/plugin install recursive-research

Option C — As a standalone skill (minimal, no plugin manager)

If you prefer to bypass the plugin ecosystem, copy only the skill markdown:

Linux / macOS:

git clone https://github.com/Anjos2/recursive-research.git
mkdir -p ~/.claude/skills/recursive-research
cp recursive-research/plugins/recursive-research/skills/recursive-research/SKILL.md ~/.claude/skills/recursive-research/

Windows (PowerShell):

git clone https://github.com/Anjos2/recursive-research.git
New-Item -ItemType Directory -Force -Path "$HOME/.claude/skills/recursive-research"
Copy-Item recursive-research/plugins/recursive-research/skills/recursive-research/SKILL.md "$HOME/.claude/skills/recursive-research/"

Invoked without namespace as /recursive-research. Verify with /help inside Claude Code.

Usage

Standard invocation

/recursive-research:recursive-research     # if installed as plugin
/recursive-research                        # if installed as standalone skill

The skill guides you interactively. Answer in natural language.

Resume a paused research

/recursive-research:recursive-research --resume <slug>

<slug> = kebab-case name of the topic (e.g., episodic-memory-humans).

List saved research

/recursive-research:recursive-research --list

Usage examples by domain

"PhD level" criterion

The skill only declares PhD when all 5 criteria are met:

1. Coverage ≥80% across all seed threads
2. ≥3 Tier-1 sources per thread
3. New-finding saturation ≤5% for 3 consecutive cycles
4. Munger inversion applied to the knowledge (what I don't know, what sources contradict, what biases exist)
5. ≥3 explicit cross-thread connections between different threads

If any criterion fails, it does not declare PhD and keeps iterating — or asks for confirmation upon hitting the cycle cap (default 20, configurable).

Source tiering

Every consulted source is logged in a per-tier file for later audit.

Pre-loaded seed sources

The skill auto-suggests reliable sources by domain. Examples:

• Science: arXiv, Semantic Scholar, Google Scholar, Connected Papers, OpenReview
• Medicine: PubMed, Cochrane Library, WHO, ClinicalTrials.gov
• Humanities: JSTOR, SSRN, Project MUSE
• Code: GitHub, Context7, RFCs, W3C specs
• Data: World Bank, OECD Data, Our World in Data, Pew Research
• Art/culture: Europeana, Google Arts & Culture, Internet Archive, Project Gutenberg

The user can add or reject any before starting.

WDM + Munger inversion

Decision frameworks applied at each non-trivial autonomous step:

• WDM (Weighted Decision Matrix) — enumerate 3+ viable alternatives, criteria with weights, 1-5 scoring, compared totals.
• Munger inversion (Charlie Munger via Jacobi) — ask inverted about the winning option: "how would it fail? what bias does it have? what am I ignoring?"

The skill applies both to:

• Select seed threads
• Select sources per cycle
• Decide when to close the research
• Validate the consolidated knowledge at the end

Reference: Charlie Munger's essay on mental inversion.

Recommended MCPs

The skill detects what you have and uses them in this order:

1. Firecrawl MCP (highly recommended) — AI-optimized scraping
2. Context7 MCP — official library docs
3. WebSearch + WebFetch (built-in Claude Code) — universal fallback
4. Chrome DevTools MCP — only when content requires real JS execution; generally too slow for large-scale research

Generated files

In memoria/investigaciones/<slug>/ of the active project:

• estado.md · hilos.md · hallazgos.md
• fuentes-tier-1.md · fuentes-tier-2.md · fuentes-tier-3.md · fuentes-rechazadas.md
• ciclo-01.md, ciclo-02.md, ..., ciclo-N.md (checkpoints)
• sintesis.md · acciones.md · gaps.md (upon closing)

If memoria/ doesn't exist in the project, the skill creates it (notifying the user) — it's an explicit dependency.

Repository structure

recursive-research/                              ← the repo is a Claude Code marketplace
├── .claude-plugin/
│   └── marketplace.json                        ← marketplace manifest (declares the plugin)
├── plugins/
│   └── recursive-research/                     ← the plugin (named same as the marketplace)
│       ├── .claude-plugin/
│       │   └── plugin.json                     ← plugin manifest
│       └── skills/
│           └── recursive-research/
│               └── SKILL.md                    ← the skill instructions (the actual content)
├── LICENSE                                      ← MIT
├── README.md                                    ← this file
├── PRIVACY.md                                   ← privacy policy (no data collection)
└── .gitignore

Contributing

Issues and PRs welcome. If you improve a criterion, add a tier, find a new anti-pattern, or want support for more domains: open a PR.

Roadmap

• Native integration with reference managers (Zotero, Mendeley)
• Export research to academic formats (LaTeX, BibTeX)
• Collaborative mode — multiple agents investigating threads in parallel
• Automatic source-quality metrics (h-index, journal impact factor)
• Support for PDFs behind legal paywalls

GitHub topics

claude-code · claude-code-skill · claude-code-plugin · ai-agent · research-tool · recursive-research · knowledge-management · weighted-decision-matrix · mental-models

Privacy

recursive-research does not collect any user data. Everything runs locally on your machine. See PRIVACY.md for the full policy.

License

MIT — use, modify, distribute freely. Just keep the copyright notice.

Acknowledgments

• Charlie Munger — for "Invert, always invert" (via Carl Jacobi)
• Claude Code team — for the skill and plugin format
• Anthropic — for the model that makes this possible

If this skill was useful, consider giving the repo a ⭐.