research-reviewer

🧪 Research Reviewer

An iterative, multi-agent academic peer-review pipeline for Claude Code. It respawns a brutally honest "fresh pair of eyes" Reviewer Agent every round, scores your paper out of 100, refactors the weak spots, and loops until the manuscript hits a flawless 100/100.

---

Why?

Self-review is biased. After staring at your own thesis for weeks you stop seeing the missing error bars, the undefined symbol in Eq. 7, the overlapping legend in Figure 4, and the claim of "universal superiority" that three datasets don't support.

Research Reviewer fixes that by treating every review as a blind, first-time review. Each iteration spawns a brand-new Reviewer Agent with zero memory of previous drafts or scores — so it can't get complacent, can't overfit to the last version, and keeps hunting for reasons to reject. You apply the fixes, it re-reviews from scratch, and the score climbs.

How it works

            ┌──────────────────────────────────────────────┐
            │   Manuscript (.tex / .md / .pdf + figures)    │
            └──────────────────────┬───────────────────────┘
                                   │
                                   ▼
              ╭───────── fresh, blind subagent ─────────╮
              │   🧐  Reviewer Agent  (no prior memory)  │
              │   • scores X / 100                       │
              │   • itemized deductions → 4 pillars      │
              │   • refactored, polished text            │
              │   • "Required Experiments" checklist     │
              ╰─────────────────────┬───────────────────╯
                                    │ apply fixes
                                    ▼
                         score == 100/100 ?  ──No──┐
                                    │ Yes          │  (respawn fresh eyes)
                                    ▼              └──────────┘
                              ✅  Done

A new agent every loop = no cognitive bias, no "I already approved this," no overfitting.

The four pillars it grades on

Pillar	What it hunts for
1. Methodological Soundness & Mathematical Precision	Unstated assumptions, undefined variables, inconsistent notation/dimensions, arbitrary/unjustified design choices.
2. Empirical Rigor & Validity of Claims	Missing error bars / CIs / p-values, unfair or under-tuned baselines, ablations that change >1 variable at once.
3. Scope of Validity & Narrative Honesty	Overclaiming beyond the data, no honest limitations / failure modes / scalability discussion.
4. Visual Layout & Editorial Integrity	Fails the 30-second graphic test, overlapping labels/legends, orphan headings, hallucinated or placeholder citations.

Install

Option A — Plugin marketplace (recommended, one line)

This repo is a Claude Code plugin marketplace. Inside Claude Code:

/plugin marketplace add roeimichael/Research-Reviewer
/plugin install research-reviewer@research-reviewer

That's it — the research-reviewer skill is now available in every session. Update later with /plugin marketplace update research-reviewer, remove with /plugin uninstall research-reviewer@research-reviewer.

Option B — Manual copy

Drop the skill straight into a skills/ folder Claude Code already reads:

git clone https://github.com/roeimichael/Research-Reviewer.git
cp -r Research-Reviewer/skills/research-reviewer ~/.claude/skills/   # global, all projects
# or:  cp -r Research-Reviewer/skills/research-reviewer .claude/skills/   # this project only

On Windows (PowerShell):

git clone https://github.com/roeimichael/Research-Reviewer.git
Copy-Item -Recurse Research-Reviewer\skills\research-reviewer ~\.claude\skills\

Usage

In Claude Code, just ask — the skill auto-triggers on review intent:

> review my thesis in ./paper/main.tex and get it to 100/100

or invoke it explicitly:

> /research-reviewer

Claude will:

1. Read your manuscript (and figures/tables/datasets if you point to them).
2. Spawn a fresh Reviewer Agent that scores it X/100 with itemized deductions.
3. Apply the refactored text and notation fixes back into your paper.
4. Hand you a Required Experiments checklist for the things only you can run.
5. Respawn a new blind reviewer and repeat until 100/100.

Customize