PaperOrchestra
A pluggable skill pack that implements the PaperOrchestra multi-agent research-paper-writing pipeline for any coding agent — no API keys, no LLM SDKs, just instruction documents and deterministic helpers.
Last reviewed by the paperbanana team on Jul 13, 2026
Install
git clone <this repo> ~/paper-orchestra
cd ~/paper-orchestra
pip install -r requirements.txt
mkdir -p ~/.claude/skills
for s in paper-orchestra outline-agent plotting-agent literature-review-agent section-writing-agent content-refinement-agent paper-writing-bench paper-autoraters agent-research-aggregator; do
ln -sf ~/paper-orchestra/skills/$s ~/.claude/skills/$s
doneWhat is PaperOrchestra?
PaperOrchestra is a skill pack that turns the prompts, schemas, halt rules, and verification pipelines from the arXiv paper "PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing" (Song et al., arXiv:2604.05018) into a set of host-agent-executable skills. It works with any coding agent that can run skills — Claude Code, Cursor, Antigravity, Cline, Aider, OpenCode — because the skills are instruction documents plus deterministic local helpers rather than an app with its own API keys or LLM SDK calls. Your coding agent does all the reasoning and web search using its own tools.
The pipeline follows the paper's five-agent design — Outline, Plotting, Literature Review, Section Writing, and Content Refinement — which the authors report beats single-agent and tree-search baselines on the PaperWritingBench benchmark by 50-68% on literature review quality and 14-38% on overall quality. PaperOrchestra ships this as seven skills total: the orchestrator (paper-orchestra) plus outline-agent, plotting-agent, literature-review-agent, section-writing-agent, content-refinement-agent, paper-writing-bench, and paper-autoraters, with the plotting and literature-review stages running in parallel.
An optional eighth skill, agent-research-aggregator, bridges the gap between scattered coding-agent history (Claude Code memory, Cursor chat logs, Antigravity worker logs, OpenClaw sessions) and the structured idea.md / experimental_log.md inputs PaperOrchestra expects. If you have been running experiments through an agent but never wrote a clean experiment log, this skill discovers, extracts, and synthesizes one for you before handing off to the main pipeline. The whole pipeline requires zero API keys under a host with native web search, and optionally integrates Semantic Scholar, PaperBanana (for figure generation), and Exa for literature discovery.
Core capabilities
Seven-skill paper-writing pipeline
Outline-agent (idea + log + template + guidelines → structured outline JSON), plotting-agent (executes the plotting plan with a VLM-critique refinement loop), literature-review-agent, section-writing-agent, and content-refinement-agent implement the paper's five pipeline steps end to end.
Citation-verified literature review
literature-review-agent web-searches candidates, verifies them against Semantic Scholar with Levenshtein similarity > 70 plus dedup, and drafts the Intro and Related Work sections with at least 90% citation integration.
Simulated peer review with halt rules
content-refinement-agent runs a simulated peer review and accepts or reverts changes per strict halt rules, with safety constraints designed to prevent gaming the evaluator.
Optional agent-log aggregator
agent-research-aggregator scans .claude/, .cursor/, .antigravity/, and .openclaw/ caches across a four-phase discovery → extraction → synthesis → formatting pipeline to produce idea.md and experimental_log.md from unwritten experiment history.
Benchmark and autorater tooling
paper-writing-bench reverse-engineers raw materials (Sparse/Dense idea, experimental log) from an existing paper to build benchmark cases; paper-autoraters runs the paper's own Citation F1, 6-axis literature-review quality, and side-by-side quality autoraters.
No API keys, no LLM SDK
Every skill is an instruction document plus purely deterministic local helpers (JSON schema validation, Levenshtein fuzzy matching, BibTeX formatting, dedup, LaTeX sanity checks). All LLM reasoning is delegated to the host coding agent.
What you can use it for
Turning structured inputs into a submission-ready paper
Drop idea.md, experimental_log.md, template.tex, and conference_guidelines.md into workspace/inputs/, then ask the agent to run the paper-orchestra pipeline on that workspace.
Writing a paper from scattered experiment history
Point the agent at a project folder or agent cache directory — "Write a paper from my work in ~/my-project" — and agent-research-aggregator extracts an idea and experimental log automatically before the pipeline runs.
Generating publication-grade figures during drafting
Set PAPERBANANA_PATH so the plotting-agent uses the PaperBanana Retriever → Planner → Stylist → Visualizer → Critic loop for diagram figures, falling back to matplotlib when unset.
Building a benchmark case from an existing paper
Use paper-writing-bench to reverse-engineer a Sparse/Dense idea and experimental log from a published paper, useful for evaluating your own writing pipeline.
Auditing a draft with the paper's own autoraters
Run paper-autoraters to score Citation F1, literature-review quality on six axes, and side-by-side paper/lit-review quality before submitting.
How to get started
- 1
Clone and install
Run git clone <this repo> ~/paper-orchestra, cd in, and pip install -r requirements.txt (deterministic helpers only, no LLM SDKs).
- 2
Symlink the skills into your host
For Claude Code, mkdir -p ~/.claude/skills and symlink each of the seven skills plus agent-research-aggregator with ln -sf. Other hosts (Cursor, Antigravity, Cline, Aider) follow the per-host instructions in skills/paper-orchestra/references/host-integration.md.
- 3
Scaffold a workspace and supply inputs
Run python skills/paper-orchestra/scripts/init_workspace.py --out workspace/, then place idea.md, experimental_log.md, template.tex, and conference_guidelines.md into workspace/inputs/ — or let agent-research-aggregator generate the first two automatically if your research is only captured in scattered logs.
- 4
Run the pipeline and optionally add integrations
Tell your coding agent "Run the paper-orchestra pipeline on ./workspace." Optionally export SEMANTIC_SCHOLAR_API_KEY, PAPERBANANA_PATH, or EXA_API_KEY to improve citation verification, figure quality, and literature discovery.
How it compares to similar skills
PaperOrchestra is a faithful, no-API-key implementation of one specific published multi-agent architecture, and works across multiple coding-agent hosts rather than being Claude Code-only. Other skills in this directory take different tradeoffs on scope and portability.
Academic Research Skills
Pick Academic Research Skills for a broader Claude Code-native pipeline with its own citation-integrity gates and reviewer calibration, rather than a reproduction of one specific paper architecture.
AI Research Feedback
Pick AI-research-feedback if you already have a drafted LaTeX paper and only need referee-style review, not a full drafting pipeline.
Paper Craft Skills
Pick paper-craft-skills if your goal is turning a finished paper into figures, slide decks, or explainer articles rather than writing the paper itself.