ARIS is a collection of Markdown-based skills that define a workflow for autonomous machine-learning research, including idea discovery, experiment automation, and review loops. Researchers and AI coding agents use it across tools such as Claude Code, Codex, Cursor, and OpenClaw without depending on a single framework. The catalogue entries are ARIS workflow skills and agents.
Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add wanshuiyin/Auto-claude-code-research-in-sleep --skill meta-applygit clone --depth 1 https://github.com/wanshuiyin/Auto-claude-code-research-in-sleepWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/wanshuiyin/auto-claude-code-research-in-sleep/meta-apply)<a href="https://agentmods.dev/skills/wanshuiyin/auto-claude-code-research-in-sleep/meta-apply"><img src="https://agentmods.dev/badge/skills/wanshuiyin/auto-claude-code-research-in-sleep/meta-apply/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/wanshuiyin/auto-claude-code-research-in-sleep/meta-apply"><img src="https://agentmods.dev/badge/skills/wanshuiyin/auto-claude-code-research-in-sleep/meta-apply.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector pass
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00081 | $0.02012 |
| Opus 5 | $0.00041 | $0.01006 |
| Sonnet 5 | $0.00016 | $0.00402 |
| Haiku 4.5 | $0.00008 | $0.00201 |
Grade A, and why
meta-apply scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 4d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 142 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Meta-Apply: the privileged landing gate for self-modification patches
This skill exists to enforce a privilege boundary. Producers like
/meta-optimize (and, later, corpus-audit) are
read-only — no Write/Edit, no apply step; they can only stage candidate patches
under .aris/meta/pending/. This skill is the only place a staged patch becomes a real
change to the corpus. Splitting "propose" from "land" across two skills with different
tool grants is what makes "a loop cannot apply its own patch" structural rather than a
sentence the producer is asked to obey.
It is human-invoked only. It runs when the user explicitly types /meta-apply after
reading the producer's REPORT. That invocation IS the landing authorization (the locked
2026-05-30 decision: the human stays at the landing gate).
The acquittal is generated HERE, not relayed
The earlier design trusted a jury_verdict: pass field in the producer's manifest. That
is forgeable — the producer writes that field, so a buggy/misaligned producer could
stage a killed patch labelled "pass." So this skill does not trust any producer-written
verdict. It runs the cross-model jury itself, at landing time, on the actual staged
diff (fresh thread, paths-only). The verdict is produced where it is consumed, by the
privileged human-invoked skill — nothing to forge.
The non-negotiable rules (jury-at-landing, reject-default)
For each staged patch the user asks to land, in order — any failure ⇒ skip & report, never silently apply:
- The human named THIS patch. Apply only patches the user listed (
/meta-apply 1,3orall); default to applying nothing. - Fresh cross-model jury PASS, obtained now. Run
mcp__codex__codex(fresh thread, NOT codex-reply;model: gpt-6-astra,config: {"model_reasoning_effort": "ultra"},sandbox: read-only, paths-only perreviewer-independence.md) on the staged.diff+ its target. Ask: does this change improve the harness without regressions; PASS or KILL + one-line reason. Include the scope-limits block fromreview-scope-limits.mdin that prompt: this jury judges ARIS's own mechanism, so an over-defensive KILL permanently blocks a good patch. Note the block bans proposing new hash binding — it is not a reason to KILL a patch that touches the existing provenance stamp. KILL ⇒ refuse. The human cannot override a KILL — they may only pick among jury-PASSED survivors. (A loop can DRIVE; only the cross-model jury can ACQUIT.) - Author ≠ reviewer family. The author is the producer's executor model; the reviewer
is the codex model that just judged it. Run
provenance.py assert_cross_family— if it raises (same family / unknown), refuse. (Here it always holds: producer=Claude, jury=codex. The check is the structural backstop.)
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 4d ago Changed 0760749e7a26
- 11d ago First seen · 142 lines · 81 tokens per session scan A 5122453add96
meta-apply is a skill published in the GitHub repository wanshuiyin/Auto-claude-code-research-in-sleep (15,970 stars, last pushed 2d ago), licensed MIT. It adds 81 tokens to every session and 2,012 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
figure-style
Publication-grade correctness and legibility rules for final-deliverable scientific figures, not exploratory plots. Use for a figure that will ship in a report, paper, export, or kept artifact. Covers data fidelity, label economy, color threading, chart choice, layout, and render-then-verify QA without imposing a…
remote-compute-ssh
Evaluate and use SSH Remote Compute before choosing where to run GPU, high-memory, parallel, batch, model-inference, bioinformatics, or other long-running scientific work; supports short remote commands and asynchronous jobs with automatic harvest and analysis.
ligandmpnn
Inverse-fold a backbone with ligand, nucleic-acid, and metal context using LigandMPNN (Dauparas et al. 2023, github.com/dauparas/LigandMPNN). Reach for this skill to redesign the residues lining a binding pocket around a bound small molecule or cofactor, to design metal-coordinating sites where the geometry must be…
evo2
Score, embed, and generate DNA sequences with Evo 2, a long-context genomic foundation model. Use this skill when: (1) Computing per-nucleotide or per-sequence likelihoods for variant effect scoring, (2) Embedding genomic windows for downstream classification, (3) Generating DNA conditioned on a prefix, (4) Scoring…
memory
Cross-project research memory. Deep-dive past projects' notes, record corrections, and save cross-project insights across all Luxas research projects.
ml-training-recipes
Battle-tested PyTorch training recipes for all domains — LLMs, vision, diffusion, medical imaging, protein/drug discovery, spatial omics, genomics. Covers training loops, optimizer selection (AdamW, Muon), LR scheduling, mixed precision, debugging, and systematic experimentation. Use when training or fine-tuning…