Skills
This directory contains the skill library for Nano-scientist. Each subdirectory is a self-contained skill that the agent can invoke during a research run.
How It Works
Skills are lazy-loaded at runtime. The agent reads skills/skills.json at startup for routing (id + description only), then loads the full SKILL.md body only when a skill is selected for execution.
Skills with allowed-tools: Bash get a real bash tool-calling loop — the model drives shell execution, sees stdout/stderr, and retries on error (up to MAX_TOOL_ROUNDS, default 16).
Skills without allowed-tools use a plain LLM call with no tool access.
Available Skills
87 skills total. Skills inherit from ARIS (Auto-claude-code-research-in-sleep). Only Bash tool access is honoured by nano-scientist; other allowed-tools values (Agent, Skill, mcp__codex__*, etc.) are present in the SKILL.md body but are not executed.
Literature & Ideation
| Skill | Description |
|---|
arxiv | Search, download, and summarize arXiv papers |
alphaxiv | Quick single-paper lookup via AlphaXiv LLM-optimized summaries |
deepxiv | Layered paper access and section-level reading via DeepXiv |
exa-search | AI-powered web search via Exa with content extraction |
gemini-search | Research paper search via Gemini for broad literature discovery |
openalex | Academic papers via OpenAlex API: citation data, affiliations, funding |
paper-navigator | Find and read papers: search, citation traversal, arXiv monitoring, SOTA |
research-lit | Search and analyze research papers, find related work |
research-survey | Structured literature survey reports: outline, draft, section expansion, assembly |
semantic-scholar | Published venue papers via Semantic Scholar: citation counts, venue metadata |
comm-lit-review | Communications-domain literature review |
novelty-check | Verify research idea novelty against recent literature |
Idea Generation & Refinement
| Skill | Description |
|---|
idea-creator | Generate and rank research ideas from a broad direction |
idea-discovery | Full idea discovery pipeline: research-lit → idea-creator → novelty-check → review |
idea-discovery-robot | Idea discovery adapted for robotics and embodied AI |
research-ideation | End-to-end ideation: literature → multi-persona generation → ELO → proposal |
research-refine | Turn vague direction into focused, implementation-oriented method plan |
research-refine-pipeline | Chain research-refine + experiment-plan in one shot |
research-review | Deep critical review of research ideas via external LLM |
kill-argument | Two-thread adversarial review: strongest rejection + point-by-point defense |
Experiment Execution
| Skill | Description |
|---|
experiment-plan | Detailed, claim-driven experiment roadmap from a refined proposal |
experiment-pipeline | Structured 4-stage execution: baseline, tuning, validation, ablation |
experiment-craft | Debug/diagnose existing experiments with structured iteration logging |
experiment-iterative-coder | Iterative code refinement via plan→code→evaluate→refine cycles |
experiment-bridge | W1.5: Read EXPERIMENT_PLAN.md, implement code, deploy, collect results |
experiment-audit | Audit experiment integrity: check for fake ground truth, phantom results |
experiment-queue | SSH job queue for multi-seed/multi-config sweeps with OOM-aware retry |
dse-loop | Autonomous design space exploration for architecture/EDA |
run-experiment | Deploy and run ML experiments on local/remote/Vast.ai/Modal GPU |
serverless-modal | Run GPU workloads on Modal serverless |
vast-gpu | Rent and manage GPU instances on vast.ai |
monitor-experiment | Monitor running experiments and collect results |
training-check | Periodically check WandB metrics to catch NaN/divergence early |
qzcli | Manage GPU jobs on the Qizhi platform via qzcli |
analyze-results | Analyze ML results, compute statistics, generate comparison tables |
result-to-claim | Judge what claims experiment results support; route to pivot/supplement/confirm |
ablation-planner | Design ablation studies from a reviewer's perspective |
system-profile | Profile scripts, processes, GPU, memory, interconnect |
Memory & Knowledge
| Skill | Description |
|---|
evo-memory | Persistent research memory: M_I + M_E stores via IDE/IVE/ESE mechanisms |
research-wiki | Persistent knowledge base: papers, ideas, experiments, claims, relationships |
Paper Writing
| Skill | Description |
|---|
paper-planning | Pre-writing planning: story design, experiment planning, figure design, timeline |
paper-plan | Generate a structured paper outline from review conclusions and results |
paper-write | Draft LaTeX paper section by section from an outline |
paper-writing | Full paper writing with 11-step workflow and LaTeX templates |
paper-review | Self-review before submission: adversarial stress-testing, 5-aspect checklist |
paper-rebuttal | Peer-review rebuttal: score diagnosis, comment prioritization, champion strategy |
rebuttal | W4: Full submission rebuttal pipeline under venue limits |
paper-figure | Generate publication-quality figures and tables from experiment results |
figure-spec | Deterministic SVG diagrams from structured JSON (FigureSpec) |
mermaid-diagram | Generate Mermaid diagrams (flowcharts, sequence, ER, Gantt, etc.) |
paper-illustration | AI illustrations for papers via Gemini image generation |
paper-illustration-image2 | AI illustrations via Codex native image generation (experimental) |
figure-description | Generate formal patent drawing descriptions from figures |
paper-compile | Compile LaTeX to PDF with error fixing |
paper-slides | Conference slides: Beamer LaTeX → PDF + PPTX with speaker notes |
paper-talk | End-to-end conference talk: paper → slides → assurance checks → export |
paper-poster | Conference poster: LaTeX → A0/A1 PDF + PPTX + SVG |
slides-polish | Per-page Codex review + python-pptx/Beamer fixes for talk slides |
formula-derivation | Structure and derive research formulas into paper-ready derivations |
proof-writer | Write rigorous mathematical proofs for ML/AI theory |
proof-checker | Verify and fix mathematical proofs via cross-model review |
citation-audit | Zero-context verification of bibliographic entries |
paper-claim-audit | Verify paper numbers/claims against raw result files |
research-pipeline | Full pipeline: idea discovery → implementation → review loop → paper |
auto-review-loop | Autonomous multi-round review loop: review → fix → re-review |
auto-review-loop-llm | Review loop using any OpenAI-compatible LLM API |
auto-review-loop-minimax | Review loop using MiniMax API |
auto-paper-improvement-loop | Autonomously improve a paper: review → fix → recompile (2 rounds) |
resubmit-pipeline | Resubmit polished paper to a different venue under hard constraints |
overleaf-sync | Two-way sync between local paper directory and Overleaf |
study-workflow | Research workflow diagram as PNG via gpt-image-2 |
writing-systems-papers | Structural blueprint for OSDI/SOSP/ASPLOS/NSDI/EuroSys papers |
Patent
| Skill | Description |
|---|
pixel-art | Generate pixel art SVG illustrations for READMEs/docs/slides |
feishu-notify | Send notifications to Feishu/Lark (webhook or interactive) |
grant-proposal | Draft grant proposals (KAKENHI, NSF, NSFC, ERC, etc.) |
meta-optimize | Analyze usage logs, propose optimizations to SKILL.md files |
Skill Anatomy
my-skill/
SKILL.md # required — frontmatter + body
scripts/ # optional — executable scripts invoked via bash tool
references/ # optional — docs loaded into agent context
SKILL.md Frontmatter
---
name: my-skill
description: "One-line summary used for routing."
allowed-tools: Bash # grants bash tool-calling loop with error feedback
required-keys: [MY_API_KEY] # optional; skill filtered out at startup if key missing
metadata:
author: YourName
version: '1.0.0'
tags: [relevant, keywords]
---
Adding a Skill
- Create
skills/my-skill/SKILL.md with the frontmatter above and your instructions in the body.
- Add an entry to
skills/skills.json:
{ "id": "my-skill", "description": "One-line description shown to the agent." }
- Validate:
python skills/validate_skills.py
Validating Skills
python skills/validate_skills.py
Checks all skills/*/SKILL.md files for required frontmatter fields, valid allowed-tools, and matching name vs directory name.