A large markdown skill set for autonomous ML research: idea discovery, experiment runs and cross-model review loops where a second model critiques the first.
A deep, actively maintained research workflow, but it depends on a second model, can launch paid compute, logs your prompts locally through hooks, and is large to adopt.
ARIS is a methodology packaged as roughly 200 skills (the file metrics count 197, averaging 16 KB; the largest, auto-review-loop, is about 75 KB). The central idea is that one model executes while a different model reviews: the executor (Claude Code by default) writes ideas, code, papers and fixes, and an independent reviewer (Codex through an MCP bridge by default, or MiniMax, ModelScope-hosted or other models) scores the work and sends back criticism for another round, up to a maximum number of rounds. Skills cover literature search (arXiv, alphaXiv, DeepXiv), idea generation and ablation planning, experiment bridging and a job queue, result analysis, claims drafting, citation auditing and cross-checking, paper-improvement loops, figure specs, slides and talks. Constants such as CODE_REVIEW and AUTO_DEPLOY control whether reviewed code is launched automatically, and the skills mention Vast.ai and Modal GPU lifecycles and cost checks.
A meta-optimize workflow uses Claude Code hooks to log skill calls, tool failures and your prompts to local events.jsonl files, then proposes reviewed patches to the skills, applied only with your consent. A few Python scripts back the queue, citation checks and figure rendering. The project also ships a standalone CLI and plugins, and the README lists adapters for Codex, Cursor, Copilot CLI and OpenClaw.
197 skills averaging 16 KB with long, rule-heavy workflows, a few Python helpers for the queue, citations and figures, and an MCP bridge, but no evals in the repo.
Counted from the repository file tree. How we score depth.
This skill asks for more access than most, or sends data to another service. That is usually a design choice rather than bad intent, but you should know before you install it.
A large research-automation skill suite (literature search, auto review loops, paper and figure tooling, GPU experiment queue) in which almost every skill pre-approves unrestricted Bash. Review loops send your drafts and code to a second model provider, the experiment skills can launch jobs on remote machines over SSH (auto-deploy is on by default in experiment-bridge), and install steps write to your global Claude config and install optional packages unpinned. Local hook logging is opt-in by copying a template; no remote telemetry was found.
The repository (761 files) on Oct 29, 2026. Not a full review: 699 of 761 files were not downloaded (large repository; we reviewed the skill folders); we read 2 of 4 code files; it relies on code outside the repository: Codex CLI, Gemini CLI, Exa, DeepXiv, Firecrawl and other packages/services the skills install or call; MCP server code, tools/, templates/ (including claude-hooks/meta_logging.json) and the ARIS runtime described in the README; these paths are not in the local copy (about 700 readable files missing); skills/citation-crosscheck/scripts/cx_verify.py (84 KB) and skills/figure-spec/scripts/figure_renderer.py were only searched by keyword (no network or exec calls beyond rsvg-convert/cairosvg found), not read in full.
NVIDIA SkillSpector static scan: 172 pattern hits (Agent Snooping 106, Data Exfiltration 20, Excessive Agency 17, Rogue Agent 8, Tool Misuse 5, Prompt Injection 3). The 172 hits are mostly skill cross-references and ~/.claude paths where API keys are read, an uninstall one-liner that removes only copied skill folders, 'git reset --hard' and 'rm -rf' listed as prohibited in dse-loop, and the shell=True in queue_manager.py (it runs manifest commands by design). The mermaid-diagram 'hidden instructions' hit is a zero-width space placed before nested code fences in a Markdown example, not an instruction.
A review, not a certification. Skills change after we check them — read the files before installing.
MIT — commercial use allowed.
Yes. Review loops have a maximum round count, experiments go through a queue manager, and the README is built around unattended runs, though auto-deploy and paid GPU use should be bounded by you.
Five very detailed Claude Code skills for the academic pipeline: deep research, paper writing, peer review, revision and systematic-review screening.
98 skills for AI research and ML engineering, plus an autoresearch orchestrator that runs a two-loop experiment and synthesis cycle.
Academic research plugin with 15 skills; literature reviews check citations and quotes against OpenAlex, arXiv and Europe PMC.
Eleven skills that reproduce deep-learning repos from their README with a reviewed plan, bounded run and evidence bundle.
ARIS is an open-source set of markdown skills for autonomous machine-learning research. A main agent does the work and a different model reviews it in loops, covering idea discovery, experiments, result analysis, citation checks and paper revision.
Clone the repo and copy its skills folder into ~/.claude/skills (global) or a project's skills folder, as the README shows, then register the Codex review bridge from Quick Start. It is also offered as Claude Code and Codex plugins and as a standalone CLI.
That is its design goal: review loops stop at a maximum number of rounds and experiments run through a queue. You should still cap rounds, spending and AUTO_DEPLOY behavior yourself, since it can start jobs and cloud instances.
We found no hidden remote calls, but it is a broad workflow: it can run shell commands for experiments, send content to a second model provider, and hooks log your prompts locally. Review the settings before leaving it unattended.