An LLM-driven research assistance framework for Claude Code.
Copyright (c) 2026 Oğuz Gençer · Licensed under CC BY-NC 4.0 · See POSITIONING.md for allowed/disallowed uses.
A modular suite of Claude Code skills, slash commands, and sub-agents that assist a human researcher through the literature → manuscript → review → revision lifecycle. The framework provides:
- Structured workflows for common scholarly tasks (lit review, fact verification, systematic review, manuscript drafting, peer review)
- Sub-agent personas for division of cognitive labor (question formulation, source discovery, synthesis, etc.)
- Templates and references anchored to public-domain academic methodologies (PRISMA 2020, APA 7, IMRaD, Socratic method)
- Checkpoint discipline — human approval is required at key decision points; the framework will not run unattended on substantive judgments
This is assistive, not autonomous. It will not write your paper for you. It handles tedious mechanics (search structuring, citation formatting, consistency checking, draft assembly) and leaves the intellectual work — what to argue, what to measure, how to interpret — to you.
researcher_agent is built for atomic artefact production — taking
a research question to a verified, peer-review-grade deliverable (a
systematic review, a manuscript, a critique). Each artefact is
self-contained, carries claim-level provenance, and does not depend on
the framework remembering anything between sessions.
It is deliberately not a knowledge base. It does not accumulate domain expertise over time, maintain a personal wiki, or run agents on a schedule. If you want a system that continuously ingests sources and grows a body of domain knowledge, that is a different pattern — an agentic knowledge-vault.
The two compose well side by side: a knowledge-vault for daily
accumulation (ingesting papers, building a domain wiki) and
researcher_agent for the writing, review, and disclosure moment
when that material becomes a concrete deliverable. They use separate
directories — no cross-contamination — and the handoff between them is
intentionally manual: you decide what crosses over.
| Skill | Purpose | Modes | Phase |
|---|---|---|---|
research |
Literature investigation, source verification, synthesis | 7 | 1 (available) |
compose |
Manuscript drafting, revision, formatting | 9 | 2 (available) |
critique |
Peer-review-style manuscript critique | 6 | 3 (available) |
orchestrate |
End-to-end pipeline across the other three skills | 2 | 4 (available) |
This release (v1.0.0) is feature-complete: all four planned skills ship with all 24 modes. See CHANGELOG.md for the release history.
# Clone to a stable canonical location
git clone https://github.com/drader/researcher_agent ~/researcher_agent
# Symlink the skill into a target project's .claude/skills/
cd /path/to/project
mkdir -p .claude/skills .claude/commands
ln -s ~/researcher_agent/skills/research .claude/skills/research
for c in ~/researcher_agent/commands/*.md; do
ln -s "$c" .claude/commands/$(basename "$c")
done
# Or: install for all projects at user level
mkdir -p ~/.claude/skills
ln -s ~/researcher_agent/skills/research ~/.claude/skills/researchSee docs/SETUP.md for additional installation options (plugin marketplace, copy-based install, multi-skill setup).
After installation, in any Claude Code session inside the configured project:
You: "Help me run a quick literature scan on stochastic spiking
neural networks for low-power inference."
Claude: <picks the research skill, brief mode, asks any clarifying
questions, then produces a research brief>
Or invoke a specific mode via slash command:
/research-socratic Guided question-formulation dialogue
/research-brief Short summary (500-1500 words)
/research-full Comprehensive synthesis (3000-8000 words)
/research-systematic PRISMA 2020 systematic review
/research-verify Claim-by-claim fact-check
/research-annotate Annotated bibliography
/research-evaluate Critical review of one provided source
See skills/research/SKILL.md for the full mode specification and trigger phrases.
-
Source-grounded. Every factual claim in produced output must be traceable to a cited source. The framework refuses to fabricate citations; if a needed source cannot be located, the gap is reported rather than papered over.
-
Human-in-the-loop. Major decisions (research question scope, methodology choice, inclusion/exclusion criteria, final outputs) are presented to the user for explicit approval. The framework will not silently resolve ambiguity.
-
Limited revision loops. At most two revision passes per deliverable. Issues unresolved after the second pass are recorded under "Unresolved Issues" rather than hidden.
-
Transparency over deception. The framework helps you produce better work and document where AI assisted. It does not help you conceal AI involvement; an explicit disclosure command is provided (Phase 2).
-
Voice fidelity. Mode parameters can be set so that produced prose stays closer to your prior writing style (when example text is provided), rather than defaulting to a generic LLM register.
CC BY-NC 4.0:
- ✅ Personal academic work (PhD, papers, grant proposals)
- ✅ Teaching, classroom demonstration, training
- ✅ Non-profit research collaboration
- ❌ Commercial SaaS or paid services built on this framework
- ❌ Internal use at a for-profit organization for revenue-generating work
- ❌ Resale, repackaging, or commercial API wrapping
For commercial licensing, contact the copyright holder.
Gençer, O. (2026). researcher_agent: An LLM-driven research assistance
framework [Computer software]. CC BY-NC 4.0.
Structured metadata (BibTeX, APA, Chicago, etc.) is available in CITATION.cff. GitHub's "Cite this repository" button on the repo sidebar reads from that file.
v1.0.0 (2026-05-21) — Feature-complete. All four planned skills ship: research, compose, critique, orchestrate. 24 modes, 24 slash commands, 21 sub-agent personas, 16 reference docs, 19 templates, 6 worked examples.
See ARCHITECTURE.md for the design overview and MODE_REGISTRY.md for the canonical list of modes.