pat-review

name

pat-review

description

Use when the user wants to assess, review, critique, or sanity-check a scientific manuscript or paper before journal submission — e.g. "review my paper", "is this draft submission-ready", "run a paper assessment", "run PAT". Runs PAT's multi-agent review pipeline and reports a Submission Readiness Score with per-dimension findings. Keywords: manuscript, peer review, reviewer 2, journal, submission-ready, readiness score, PAT.

Run a PAT review

Overview

PAT (Paper Assessment Tool) runs ~31 specialized AI agents over a scientific manuscript and emits a Submission Readiness Score plus per-dimension findings (writing, statistics, methods, figures, references, positioning, reporting-guideline compliance) and a ranked action plan. This skill runs PAT on a paper and summarizes the result.

Prerequisites

• Python 3.10+.
• Install the package (provides the pat command): pip install -e . from the repo root.
• A model provider — pick one:
  • Ollama (default, local, free): Ollama running locally with a model pulled; ./run.sh handles install + model pull automatically.
  • Anthropic (cloud, faster, prompt caching): export ANTHROPIC_API_KEY=...

Quickest path

./run.sh /abs/path/to/paper.pdf

run.sh verifies Python/Ollama, creates a venv, installs the package, pulls the model, and runs a full review with --html. Reports land in reports/<model-tag>/.

Manual run

pat /abs/path/to/paper.pdf --html            # console script (after pip install -e .)
python -m pat /abs/path/to/paper.pdf --html  # module form

Accepts .pdf, .txt, .md, .tex. PDFs auto-extract figures for the vision agents.

Choosing a provider

pat paper.pdf                                    # local Ollama (default, no key)
export ANTHROPIC_API_KEY=sk-ant-...
pat paper.pdf --provider anthropic --parallel 5  # cloud Claude

Pre-flight (recommended before a full cloud run)

pat --check                  # provider connectivity
pat paper.pdf --estimate     # token/cost estimate (no model calls)
pat paper.pdf --dry-run      # show plan without calling the model
pat --list-agents            # all agent IDs

Common options

Goal	Flag
HTML report (radar chart, heatmap)	--html
Reproducibility check against code	--code-file analysis.py
Add figure images for vision agents	--figures fig1.png fig2.png
Run only some agents	--agents vsnc,intro,statistics
Skip literature search (offline/faster)	--ref-backend none (or --no-scholar)
Reference backend	--ref-backend pubmed|biorxiv|pubmed+biorxiv
Compare against a prior review (revisions)	--compare reports/review_*.md
Annotated manuscript with inline comments	--annotate
Resume after a crash	--resume (or --fresh to ignore checkpoints)
Target-journal config	--config examples/review_config.json
Output directory	--output-dir reports/v2

Run pat --help for the full flag list.

Reading the result

Reports are written to reports/review_<papername>_YYYYMMDD_HHMMSS.md (and .html with --html). When summarizing back to the user:

• Lead with the Submission Readiness Score (report header).
• Surface the orchestrator's ranked top-10 action plan.
• Per agent: note SEVERITY (ok/minor/moderate/major) and top issues; prioritize major/moderate.
• Mention consensus issues (flagged by 3+ agents) from the agreement analysis, and any reporting-guideline gaps (desk-rejection risks).

Notes

• Literature agents use PubMed/bioRxiv (free, no key); use --ref-backend none or --no-scholar if offline.
• The reproducibility agent only runs with --code-file.
• Anthropic runs cost money — prefer --estimate first on long papers.