NanoResearch Turns Claude Code into a 9-Stage Research Pipeline

Papers generated by LLMs have a credibility problem: the numbers are invented. NanoResearch, a 1k-star open-source research engine, is built against exactly that — it runs the generated code on a local GPU or a SLURM cluster, collects the real results, and writes the LaTeX paper from what actually happened, not what the model imagines.

Why This Skill Matters

NanoResearch describes itself as an end-to-end autonomous AI research engine: from a research idea to a complete paper, with every data point, table, and figure backed by an experiment that actually ran. The pipeline is nine stages deep, and the repository ships a skills/ directory of Claude Code skills — including nanoresearch-ideation, nanoresearch-planning, nanoresearch-experiment, and nanoresearch-writing — plus a set of /project:* commands that drive each stage.

The Claude Code integration mode needs no API key configuration: Claude Code itself becomes the research engine. Per the README, WebSearch replaces external APIs for literature retrieval (arXiv, Semantic Scholar, Google Scholar), Bash runs experiment code, submits SLURM jobs, and compiles LaTeX, and file read/write generates the experiment code, paper, and structured artifacts.

Installation

Clone the repository and start Claude Code inside it:

git clone https://github.com/OpenRaiser/NanoResearch.git
cd NanoResearch
claude
/project:research "Your Research Topic Here"

Real Workflow: Drive the Pipeline Stage by Stage

The full pipeline is one command — /project:research <topic> — but the same skills are exposed stage by stage:

Run /project:status after any stage to see where the workspace stands; run /project:resume to pick up from the last checkpoint — the README's FAQ says the pipeline continues from the last incomplete or failed stage. Every stage's artifacts are tracked in manifest.json, so an interrupted run picks up where it stopped instead of starting over.

The Python CLI runs the same pipeline outside Claude Code, and workspaces created in Claude Code mode are fully compatible with it — you can mix the two modes on one project.

Tips

• LaTeX compilation works better with tectonic than pdflatex; the README suggests installing it with conda install -c conda-forge tectonic.
• Architecture diagrams come out better with image models — the README recommends the Nano Banana series, called via Bash in the figure_gen stage.
• Beyond Claude Code, the README documents a project-level entrypoint for Codex that reuses the existing CLI, workspace, manifest, and PaperMode logic — no second pipeline or --mode codex flag.
• The experiment stages need real compute. If no local GPU or SLURM cluster is reachable, the execution stages have nothing to run — the engine's honesty depends on it.

When Not to Use This

NanoResearch is a computational-research pipeline first: the README documents a paper-only mode (skip_stages for the setup-through-analysis stages) and a paper-draft-assistance use case, but running real experiments is what it is built for. If your question needs only a literature summary, Stage 1 alone is the relevant part and the full pipeline is overhead. And if no GPU or cluster is available for the experiments it designs, the stages downstream of planning cannot produce real data — which is the whole point of the tool.

See the leaderboard for more skills.