Robin Scientist Skill

Robin Scientist is an agent skill for Robin-style biomedical scientific discovery workflows, based on FutureHouse's Nature paper, arXiv preprint, supplementary materials, and public Future-House/robin implementation.

It helps an AI agent preserve the important details of the Robin system while using progressive disclosure: the main SKILL.md stays small, paper details live in references/, and deterministic helpers live in scripts/.

Scope

Use this skill for:

• Robin paper analysis and implementation mapping
• biomedical therapeutic hypothesis generation
• drug repurposing workflows
• disease-to-assay-to-candidate discovery loops
• Crow / Falcon / Finch role decomposition
• lab-in-the-loop experimental iteration
• LLM-judge pairwise ranking and Bradley-Terry-Luce scoring
• dAMD / ripasudil / ROCK inhibition case analysis

Do not treat this as a general all-science automation framework without explicitly adapting the domain assumptions. The demonstrated Robin implementation is biomedical and therapeutic-discovery oriented.

Source Grounding

Primary sources:

• Nature article: A multi-agent system for automating scientific discovery, published 2026-05-19
• arXiv preprint: arXiv:2505.13400, submitted 2025-05-19
• Official implementation: Future-House/robin
• FutureHouse announcement: Demonstrating end-to-end scientific discovery with Robin

This repository is not the official Robin implementation. It is a skill layer that helps an agent understand, reproduce, scaffold, and audit Robin-style workflows.

What Robin Contributes

Robin connects four pieces into one discovery loop:

1. Literature-grounded hypothesis generation through Crow and Falcon.
2. Experimental assay and therapeutic candidate ranking through an LLM-judge tournament.
3. Finch-based analysis of experimental data such as flow cytometry and RNA-seq.
4. Iterative feedback from lab results into the next round of therapeutic hypotheses.

The key distinction is lab-in-the-loop: the AI system performs the intellectual work of proposing, ranking, interpreting, and refining hypotheses, while humans still execute physical wet-lab experiments.

Repository Layout

robin-scientist/
├── SKILL.md
├── README.md
├── LICENSE
├── references/
│   ├── source-map.md
│   ├── workflow.md
│   ├── damd-case.md
│   ├── evaluation.md
│   └── prompts-and-limits.md
└── scripts/
    ├── scaffold_robin_run.py
    ├── extract_robin_tables.py
    └── btl_rank.py

File roles:

• SKILL.md: trigger, scope, workflow, and guardrails for the agent.
• references/source-map.md: source URLs, publication dates, and verified corrections.
• references/workflow.md: public repo workflow, function entrypoints, parameters, and output structure.
• references/damd-case.md: dry AMD case details, candidate rounds, assay details, and quantitative results.
• references/evaluation.md: LLM judge, BTL ranking, Finch, BixBench-style metrics, and rubrics.
• references/prompts-and-limits.md: prompt surfaces, judging criteria, and implementation limits.
• scripts/scaffold_robin_run.py: create a small runner around the official Robin repo.
• scripts/extract_robin_tables.py: recompute selected metrics from the Nature supplementary tables workbook.
• scripts/btl_rank.py: fit a local Bradley-Terry-Luce ranking from pairwise winner/loser CSV data.

Quick Start

Install or copy this folder into the skill directory used by your agent host. For a Codex-style setup:

mkdir -p ~/.agents/skills
git clone https://github.com/TeamZaobi/robin-scientist.git ~/.agents/skills/robin-scientist

Then start a fresh agent session so the host reloads available skills.

Script Examples

Create a manuscript-style Robin runner:

python3 scripts/scaffold_robin_run.py \
  --out-dir /tmp/robin-run \
  --disease "dry age-related macular degeneration" \
  --paper-defaults

Recompute selected supplementary-table metrics:

python3 scripts/extract_robin_tables.py /path/to/41586_2026_10652_MOESM3_ESM.xlsx

Fit a BTL ranking from pairwise comparisons:

python3 scripts/btl_rank.py pairwise_results.csv

Expected CSV columns include either winner_id / loser_id or Winner ID / Loser ID.

Validation

Check script syntax:

python3 - <<'PY'
from pathlib import Path
import ast

for path in Path("scripts").glob("*.py"):
    ast.parse(path.read_text(encoding="utf-8"), filename=str(path))
    print(f"OK {path.name}")
PY

Optional smoke test:

python3 scripts/scaffold_robin_run.py \
  --out-dir /tmp/robin-scaffold-test \
  --disease "dry age-related macular degeneration" \
  --paper-defaults

Important Caveats

• Nature publication date and arXiv date differ: Nature is 2026-05-19; arXiv is 2025-05-19.
• The full DOI is 10.1038/s41586-026-10652-y.
• Finch trajectory counts vary by source: paper text discusses 10 trajectories, one RNA-seq figure references eight, and the public code currently uses PARALLEL_ANALYSIS = 5.
• Do not collapse ripasudil effect sizes into one number. The paper text, human analysis, and supplementary tables report different values under different analysis contexts.
• Do not present unverified secondary-source claims as primary paper facts.
• This skill does not provide medical advice, clinical validation, or regulatory conclusions.

Relationship To The Official Robin Repo

The official Robin implementation requires Python 3.12+, Edison platform access, LLM provider credentials, and often Docker/Jupyter. This skill does not vendor that code. It provides:

• workflow interpretation
• prompt and source routing
• local scaffolding around the official repo
• BTL and supplementary-table utilities
• agent guardrails for accurate Robin discussion

Use the official repository when you need to run Robin itself.

License

Apache-2.0. This repository is an independent skill layer and does not claim affiliation with FutureHouse.