research-superpowers

A Claude Code / OpenCode plugin that turns LLM-assisted research into a disciplined, reproducible workflow.

Inspired by obra/superpowers, adapted to the research lifecycle: idea → plan → literature → ingest → analysis → draft → peer review → publication.

Built for theology, biblical archaeology, ancient history, and digital humanities — methodology-aware (hermeneutic / quantitative / mixed) so it doesn't force quantitative pre-registration onto hermeneutic work, but enforces it where it belongs.

Why

LLM-assisted research is fast but easy to get wrong:

• Citations get hallucinated. Author–year strings drift from what the source actually says.
• Synthesis pages get written from memory of the source, not the source itself.
• The "I'll cite this properly later" moment never returns.
• Pre-registration norms from quantitative psychology get jammed into hermeneutic disciplines where the hermeneutic circle is constitutive, not a methodological failure.
• "Don't cite academia.edu as a primary source" stays as a red flag the model ignores, until a structural mechanism makes it enforceable.

research-superpowers makes the discipline structural:

• 17 skills with checklists, SOFT-GATES, and red-flag tables for every phase
• 7 subagents for context-isolated heavy lifting (ingest, acquire, draft, review, analysis, literature search, lint)
• Frontmatter schema that the linter validates and your editor live-checks
• Methodology branching — hermeneutic projects skip frozen-hypothesis pre-registration; quantitative keeps it; mixed is per sub-study
• Audit-trail SOFT-GATES that prompt for a written override reason and log it, instead of silently blocking
• Two recommended MCPs for verified citations and web source-class detection (optional; plugin works without them)

Who is this for

• Researchers in theology, biblical archaeology, ancient history, digital humanities who want LLM help without losing the discipline their field expects from a manuscript.
• Anyone running mixed-methods projects where some sub-studies need formal pre-registration and others legitimately don't.
• Teams who want a shared schema for their research wiki, with structural checks (frontmatter validation, wikilink resolution, override-rate tracking).

If you write Python ML papers with frozen RCT hypotheses, this plugin's discipline machinery is overkill (use a simpler workflow). If you write archaeology / theology / DH papers with contested sources and iterative reading, this is for you.

Installation

New here? The full docs/installation.md walks through everything step by step — prerequisites, the three install paths, verification, troubleshooting. Allow 5–10 minutes.

Cowork user or no terminal? docs/installation-cowork.md is the click-only path — just Claude, no Python, no Git. Allow 3–5 minutes.

Familiar with Claude Code plugins? Three lines:

claude                                                          # open Claude Code anywhere
/plugin marketplace add leiverkus/research-superpowers          # register the marketplace
/plugin install research-superpowers@leiverkus-research         # install the plugin

That installs and activates it. Restart Claude Code and you'll see a "Research Superpowers" notice at session start.

Other install paths

• Direct from GitHub (no marketplace registration):

  /plugin install https://github.com/leiverkus/research-superpowers
  

• Local clone (for development or customisation):

  git clone https://github.com/leiverkus/research-superpowers ~/code/research-superpowers
  

  then in Claude Code:

  /plugin install ~/code/research-superpowers
  

• OpenCode: symlink skills/ into your project's .claude/skills/ — OpenCode reads SKILL.md files natively. Details in docs/installation.md.

Prerequisites

• Claude Code 2.x or newer
• Python 3.10+ with PyYAML (python3 -m pip install --user pyyaml)
• Git

If any are missing, docs/installation.md shows how to install each.

For PDF output only (Quarto + XeLaTeX/LuaLaTeX — not needed for HTML or the wiki): the
templates ship with free, OFL-licensed fonts tuned for scholarly transcription
of Semitic languages — Latin transliteration (ḥ ṣ ṭ ʾ ʿ ā ī ū), polytonic
Greek, and native Hebrew (RTL):

# macOS (Homebrew casks)
brew install --cask font-gentium-plus font-noto-sans font-fira-code
# Ezra SIL (Hebrew) has no cask — download from https://software.sil.org/ezra/
# and copy the .ttf into ~/Library/Fonts/

Full install + usage notes (incl. the Hebrew \foreignlanguage macro) are in
the template's output/README.md.

Create your first research project

After the plugin is installed, scaffold a project from the template:

cp -r ~/.claude/plugins/cache/leiverkus-research/research-superpowers/templates/research-project-template ~/Documents/my-first-project
cd ~/Documents/my-first-project

Edit the three frontmatter lines at the top of CLAUDE.md (methodology, discipline, languages), then run claude in the project root. You're ready — head to the Quickstart for your first ingest.

Recommended (optional): MCPs for verified citations

Two MCP servers by the same author add structurally verified citations and source-class detection. The plugin works without them. See docs/recommended-mcps.md for setup.

30-second example

In a research project (scaffold one from templates/research-project-template/), point the project at the shared library (.research-library) and tell Claude Code:

Ingest the Finkelstein 2003 PDF.

The ingest-source skill triggers, reads the PDF fully, derives slug finkelstein-2003, extracts bibliographic data and entities (people, places, concepts), and writes:

• knowledge/sources/finkelstein-2003.md — structured wiki page (theses, methodology, verbatim quotes with pages, mentioned entities)
• one stub knowledge/entities/<entity>.md per new entity
• one BibTeX entry in output/bibtex/references.bib
• one log line in knowledge/_meta/log.md
• runs python scripts/lint-wiki.py — exits 0

You review the diff, accept or revise, commit. Repeat per source. After ~5 sources, synthesise. After synthesis is status: stable, draft. Every claim cites a source page; every citation resolves in BibTeX; every render is checked.

Full walkthrough: docs/tutorial.md.

Skill topology

brainstorming-research → writing-research-plan → literature-review → acquire-sources ⇄ ingest-source
   → executing-research-plan → drafting-manuscript → requesting-peer-review → finishing-a-research-project

Cross-cutting (context-triggered, no phase binding):

Subagents (agents/<name>.md) provide context isolation for heavy tasks; each declares implements: <skill-name> and never duplicates the skill's checklist.

Phase graph with hermeneutic back-edges and SOFT-GATE semantics: docs/phase-flow.md.

Documentation

Repository layout

research-superpowers/
├── README.md                     # this file
├── LICENSE                       # MIT
├── CHANGELOG.md
├── CONTRIBUTING.md
├── .gitignore
├── .claude-plugin/plugin.json    # plugin manifest
├── schema/
│   ├── knowledge-frontmatter.schema.json
│   └── README.md
├── hooks/                        # SessionStart hook injects the skill index
├── skills/                       # 17 SKILL.md files (the SOT for each workflow)
├── agents/                       # 7 thin-pointer subagents
├── templates/research-project-template/
│   ├── CLAUDE.md                 # frontmatter declares methodology + discipline
│   ├── schema/                   # mirror of plugin schema (for project-local lint)
│   ├── scripts/lint-wiki.py      # validates against the schema; reports override rate
│   └── .vscode/settings.json     # yaml.schemas → live frontmatter validation
├── examples/example-project/     # a small worked example
└── docs/                         # the docs listed above

Versioning

Semantic Versioning. At 0.x.y, minor versions may include breaking changes when they buy real architectural clarity (see CHANGELOG.md — v0.2 was a major architecture overhaul). Patch versions are additive. Once we reach 1.0.0, strict semver applies.

License

MIT, Patrick Leiverkus, 2026. See LICENSE.