Research Review

---
name: research-review
description: Run an internal research critique with likely objections, severity, and a concrete revision plan. Use when the user asks to review a paper, preprint, draft, or research artifact — their own or someone else's.
argument-hint: <artifact>
---

# Research Review

Review the AI research artifact the user names (arXiv ID, URL, local file, PDF, Markdown, etc.).

Derive a short slug from the artifact name (lowercase, hyphens, no filler words, ≤5 words). Use this slug for all files in this run.

This is an execution request, not a request to explain or implement the workflow instructions. Carry out the workflow with tools and durable files. Do not answer by describing the protocol, saying what you would do, or stopping after a plan.

Do not ask for confirmation. Briefly summarize the plan to the user and continue immediately unless the user explicitly asked to review the plan first.

## Tools (Claude Code)

- Web search: `web_search`. Fetch pages: `browser` fetch.
- Academic papers: the `arxiv` MCP tools — `discover_papers` (search), `get_paper_content` (read), `answer_pdf_queries` (Q&A on a paper's PDF), `read_files_from_github_repository` (paper code). If these tools are not visible, fall back to web_search/browser fetch on arxiv.org and record the degradation.
- PDFs: read them with the Read tool (page ranges) or `pdftotext` via Bash.
- Subagents: the `Agent` tool with `subagent_type` `researcher` / `reviewer`.

## Required artifacts

- Plan: `outputs/.plans/<slug>-review-plan.md`
- Evidence notes: `outputs/.drafts/<slug>-review-evidence.md`
- Final review: `outputs/<slug>-review.md`

## Workflow

1. Create `outputs/.plans`, `outputs/.drafts`, and `outputs`.
2. Write `outputs/.plans/<slug>-review-plan.md` with:
   - artifact identifier and source type (arXiv ID, URL, local file, PDF, Markdown, etc.)
   - review criteria: novelty, empirical rigor, baselines, reproducibility, claims validity, figures/tables, metrics, related work, writing quality
   - verification checks needed for claims, figures, reported metrics, data/code availability, and linked artifacts
3. Continue immediately. Do not end after planning.
4. Inspect the artifact:
   - For local files, read or parse the file directly.
   - For PDFs, use the Read tool with page ranges or `pdftotext`. If PDF parsing fails, use any available fallback extraction, record the failure, and still produce a blocked or partial review artifact.
   - For arXiv IDs or URLs, fetch the paper/source directly (via the alphaxiv `get_paper_content` MCP tool or browser fetch) and record the URL.
   - Inspect linked code, datasets, supplemental material, or citations when they are reachable and materially affect the review.
5. Write evidence notes to `outputs/.drafts/<slug>-review-evidence.md` before writing the final review. Include quoted/paraphrased claims, observed methods, reported metrics, baseline comparisons, reproducibility facts, and every inspected source path or URL.
6. Use `researcher` and `reviewer` subagents only when the artifact is large enough to benefit from delegation. If subagents would only add overhead, do the lead-owned review directly. Never merely say a subagent was spawned; either call subagent shells or continue yourself.
7. Write exactly one final review artifact to `outputs/<slug>-review.md` with:
   - Summary Assessment
   - Strengths
   - Critical Issues
   - Major Issues
   - Minor Issues
   - Reproducibility and Verification
   - Inline Annotations tied to sections, claims, figures, or tables where possible
   - Recommendation
   - Sources
8. If the artifact cannot be parsed or critical evidence is unavailable, still write `outputs/<slug>-review.md`. Mark the affected sections with `Verification: BLOCKED`, explain exactly what failed, and distinguish blocked checks from actual paper weaknesses.
9. Before responding, verify on disk that `outputs/<slug>-review.md` exists. If it does not exist, create it immediately as a blocked review artifact with the failure reason.

Never end with planning-only chat. Never ask what to do next. Never claim the review is complete unless `outputs/<slug>-review.md` exists.