wanshuiyin--auto-claude-code-research-in-sleep/docs/SKILLS_CATALOG.md

ARIS Skills Catalog

Every skill that ships with ARIS, grouped by role. 82 skills as of the latest update; new skills land via PR and get added to the table below.

• Each Skill link goes to the canonical SKILL.md (the LLM-readable spec).
• Role is a one-line summary — see the SKILL.md for the full contract, phases, and triggers.
• Requires lists external dependencies beyond ARIS core (Codex MCP, Gemini API, Modal account, LaTeX toolchain, etc.). None means it works out of the box on a standard install.

Codex CLI mirror: every skill below has a parallel implementation under skills/skills-codex/ for Codex CLI users. The mirror swaps the Codex-MCP reviewer path for Codex-native spawn_agent + send_input. SKILL semantics are otherwise identical; the table below tracks the main-tree canonical files.

---

🏗️ Workflow Orchestrators

End-to-end pipelines that chain many sub-skills. Most users start here.

Skill	Role	Requires
/research-pipeline	Full chain — Workflow 1 → 1.5 → 2 → 3, from research direction to submission-ready paper	Codex MCP, LaTeX, GPU
/idea-discovery	Workflow 1 — research-lit → idea-creator → novelty-check → research-review → research-refine-pipeline	Codex MCP
/idea-discovery-robot	Workflow 1 adapter for robotics / embodied AI — robotics-aware literature survey + benchmark-anchored ideation	Codex MCP
/experiment-bridge	Workflow 1.5 — read experiment plan → implement code → sanity check → deploy to GPU → collect initial results	GPU (local / remote / Vast / Modal)
/auto-review-loop	Workflow 2 — autonomous review → fix → re-review until positive or max rounds; uses Codex MCP reviewer	Codex MCP
/auto-review-loop-llm	Same as Workflow 2 but uses any OpenAI-compatible LLM via llm-chat MCP server	llm-chat MCP
/auto-review-loop-minimax	Workflow 2 variant pinned to MiniMax API	MiniMax API key
/paper-writing	Workflow 3 — paper-plan → paper-figure → illustration → paper-write → paper-compile → auto-paper-improvement-loop	Codex MCP, LaTeX
/rebuttal	Workflow 4 — parse reviews → atomize → strategy → draft → safety check → stress test → 2-version output → follow-ups	Codex MCP
/resubmit-pipeline	Workflow 5 — text-only port across venues (no new experiments, no bib edits) — isolation → anonymity → audits --soft-only → microedit → kill-argument gate → compile + push	Codex MCP, LaTeX
/paper-talk	Workflow 6 — paper → slide outline → Beamer + PPTX → per-page polish → assurance audits → final report	Codex MCP, LaTeX, python-pptx
/research-refine-pipeline	Sub-pipeline used by /idea-discovery — refine method + plan experiments in one chain	Codex MCP
/patent-pipeline	Full patent drafting — invention → claims → spec → jurisdiction format (CN / US / EP)	Codex MCP
/dse-loop	Autonomous design-space exploration loop for computer architecture / EDA — run → analyze → tune → iterate until objective met	Domain-specific tools
/meta-optimize	Workflow M — analyze ARIS usage logs and propose SKILL.md / prompt / default-parameter improvements (outer-loop self-evolution)	Codex MCP, hook logging
/meta-apply	Privileged landing gate — the only skill allowed to mutate the skill corpus; lands /meta-optimize patches the human approved, after a fresh cross-model jury PASS on the staged diff (read-only producer ≠ privileged applier)	Codex MCP, human-in-loop

📚 Literature & Search

Paper retrieval, summarization, novelty verification.

Skill	Role	Requires
/research-lit	Multi-source literature search — Zotero / Obsidian / local PDFs / web / arXiv / S2 / DeepXiv / Exa / Gemini / OpenAlex with cross-source dedup	None (sources gated by MCP / SDK availability)
/arxiv	Search, download, summarize arXiv papers; multi-result table + per-paper detail	None
/semantic-scholar	Published-venue paper search (IEEE / ACM / Springer) — citation counts, venue metadata, TLDR	None (rate-limited without S2 API key)
/deepxiv	Progressive paper reading — search → brief → head → section → trending → web search	pip install deepxiv-sdk
/exa-search	AI-powered broad web search with content extraction — blogs, docs, news, papers	pip install exa-py + EXA_API_KEY
/web-debug-search	Multi-source debugging search across GitHub, Stack Exchange, Chinese technical communities, and general web — routing, compatibility, and discovery-only results	None
/openalex	OpenAlex API search — 250M+ open citation graph, institutional affiliations, funding data	pip install requests
/gemini-search	Gemini-driven literature discovery — decomposes topics into sub-problems, aliases, variants	gemini-cli v0.40+
/alphaxiv	Quick single-paper lookup via AlphaXiv — three-tier fallback (overview → markdown → LaTeX source)	None
/comm-lit-review	Communications-domain literature review with Claude-style knowledge-base-first retrieval — wireless / networking / satellite / Wi-Fi / cellular	None
/novelty-check	Verify a research idea is novel against recent literature — multi-source search + cross-model verification + closest-prior-work table	Codex MCP

💡 Ideation & Method Design

Generating, refining, planning research ideas before implementation.

Skill	Role	Requires
/idea-creator	Brainstorm 8-12 ideas, filter by feasibility, pilot on GPU, rank by signal	Codex MCP, GPU for pilots
/research-refine	Iterative method refinement — problem anchor → up to 5 review rounds → score ≥ 9	Codex MCP
/experiment-plan	Turn a refined proposal into a claim-driven experiment roadmap — ablations, budgets, run order	None
/ablation-planner	Design ablation studies from a reviewer's perspective (after main results pass /result-to-claim)	Codex MCP
/formula-derivation	Structure theory derivations — organize assumptions, build derivation chains, turn scattered equations into coherent narrative	None

🧠 Proof Engineering

Orchestrating and independently auditing mathematical proofs.

Skill	Role	Requires
/proof-orchestrator	Stateful proof-run orchestration: run directories, cross-run continuation, manual GPT Pro handoff packages, optional DeepSeek second opinion (additional evidence only — /proof-writer drafts proofs, /proof-checker owns verification and submission acceptance)	None by default; llm-chat MCP only for optional DeepSeek review

🧪 Experiments & Infrastructure

GPU job submission, scheduling, monitoring, profiling.

Skill	Role	Requires
/run-experiment	Deploy experiments to local / remote / Vast.ai / Modal GPU	GPU (configurable)
/monitor-experiment	Monitor running experiments, check progress, collect results	None
/analyze-results	Compute statistics, generate comparison tables, surface insights from experiment results	None
/experiment-queue	SSH job queue for multi-seed / multi-config sweeps — OOM retry, stale-screen cleanup, wave gating, crash-safe state	SSH access
/vast-gpu	Rent, manage, destroy on-demand GPU on Vast.ai	Vast.ai account + vast-cli
/serverless-modal	Run GPU workloads on Modal — zero-config serverless, auto scale-to-zero	pip install modal + Modal account
/qzcli	Manage GPU compute jobs on the Qizhi (启智) platform via qzcli (kubectl-style CLI)	qzcli installed
/training-check	Periodically poll W&B metrics during training — catch NaN, loss divergence, idle GPUs early	W&B account
/system-profile	Profile a target (script / process / GPU / memory / interconnect) with external tools + code instrumentation; produce actionable report	Profiling tools

🛡️ Review, Audit & Assurance

Cross-model critique, integrity checking, evidence verification.

Skill	Role	Requires
/research-review	Single-round deep critical review from external LLM (Codex GPT xhigh by default; oracle-pro route for Pro tier)	Codex MCP (or Oracle MCP)
/experiment-audit	Cross-model integrity check of experiment code + results — catches fake ground truth, score-normalization fraud, phantom results, scope inflation	Codex MCP
/result-to-claim	Map experimental results to intended claims — judges what's supported, what's not, what's missing; routes to next action	Codex MCP
/paper-claim-audit	Zero-context numeric verification — every number / comparison / scope claim in the paper checked against raw result files by a fresh reviewer (no confirmation bias)	Codex MCP
/citation-audit	Bibliography audit — existence + metadata correctness + context appropriateness for every \cite{}; --soft-only mode for frozen-bib resubmits	Codex MCP, web access
/proof-checker	Rigorous mathematical proof verification — 20-category issue taxonomy, two-axis severity, side-condition checklists, counterexample red team, proof-obligation ledger	Codex MCP
/kill-argument	Two-thread adversarial review — Thread 1 writes the strongest 200-word rejection memo; Thread 2 (independent) defends point-by-point and surfaces still-unresolved issues	Codex MCP
/integrity-forensics	SHA-pinned thin launcher for Anti-Autoresearch — evidence-ledger forensic sweep (46 patterns, deterministic adjudicator) → typed BLOCK/WARN/NO_NEW_BLOCKER gate + append-only obligations ledger; default pre-submission self-audit in /paper-writing	git, Codex MCP (via upstream)

📝 Paper Writing & Figures

LaTeX generation, figure / diagram production, prose polishing.

Skill	Role	Requires
/paper-plan	Generate a structured paper outline from review conclusions + experiment results — claims-evidence matrix, section structure, figure plan, citation scaffolding	None
/paper-write	Section-by-section LaTeX generation (ICLR / NeurIPS / ICML / IEEE / ACL / AAAI / CVPR / ACM MM). Anti-hallucination BibTeX via DBLP / CrossRef	None
/paper-figure	Publication-quality matplotlib / seaborn plots + LaTeX comparison tables from experiment results	matplotlib / seaborn
/figure-spec	Deterministic JSON → SVG renderer for architecture / workflow / pipeline / audit-cascade diagrams. Shape-aware edge clipping, self-loops, CJK width estimation	None
/paper-illustration	AI architecture + method illustrations via Gemini image generation, with Claude-supervised iterative refinement	GEMINI_API_KEY
/paper-illustration-image2	Codex-native image generation alternative — uses ChatGPT Plus / Pro quota via local Codex app-server bridge (no Gemini key)	Codex app-server + codex-image2 MCP bridge
/mermaid-diagram	Generate Mermaid diagrams from requirements — flowcharts, sequence, class, ER, Gantt, with syntax verification	None
/pixel-art	Generate pixel-art SVG illustrations for READMEs, docs, slides	None
/paper-compile	Compile LaTeX paper to PDF — auto-fix errors, submission readiness checks	LaTeX (latexmk, pdfinfo)
/auto-paper-improvement-loop	2-round content review + format check — typical 4 / 10 → 8.5 / 10 score lift. --edit-whitelist mode for resubmits	Codex MCP
/proof-writer	Draft rigorous mathematical proofs for ML / AI theory — theorems, lemmas, propositions, corollaries; fill in missing steps; formalize sketches	None
/writing-systems-papers	Paragraph-level structural blueprint for 10-12 page systems papers — page allocation, paragraph templates, writing patterns for OSDI / SOSP / ASPLOS / NSDI / EuroSys	None
/grant-proposal	Structured grant proposal drafting — KAKENHI (JP), NSF (US), NSFC (CN including 面上 / 青年 / 优青 / 杰青 / 海优 / 重点), ERC (EU), DFG (DE), more	None

🎤 Talks, Posters & Resubmission

After-paper outputs and venue porting.

Skill	Role	Requires
/paper-slides	Conference presentation — Beamer LaTeX → PDF + editable PPTX + speaker notes + full talk script	LaTeX, python-pptx
/slides-polish	Per-page Codex review + targeted python-pptx / Beamer fixes (font scaling, frame resize, banner-as-tcolorbox, italic leak guard, em-dash spacing, CJK font hint, anonymity placeholder discipline)	Codex MCP, python-pptx
/paper-poster-html	Default conference poster — single HTML/CSS file with measurement-driven hard gates (two-hue tokens, real paper figures with provenance, anti-patch-loop fix vocabulary) → print-ready PDF via headless Chromium	Playwright (Chromium), PyMuPDF
/paper-poster	DEPRECATED — redirect stub to /paper-poster-html (legacy LaTeX pipeline retired; in git history)	—

(Orchestrators /paper-talk for the talk pipeline and /resubmit-pipeline for venue porting live under Workflow Orchestrators.)

📜 Patents

End-to-end patent drafting and prior-art workflow.

Skill	Role	Requires
/invention-structuring	Structure a raw invention idea into a formal invention disclosure	None
/claims-drafting	Draft patent claims — independent + dependent, with anti-pattern checks	None
/embodiment-description	Write detailed embodiment descriptions for the patent specification	None
/specification-writing	Full patent specification from claims + invention disclosure	None
/figure-description	Generate formal drawing descriptions for patent figures	None
/prior-art-search	Search patent databases + academic literature for prior art relevant to an invention	None (web access)
/patent-novelty-check	Assess patent novelty and non-obviousness against prior art (patentability evaluation)	Codex MCP
/patent-review	External patent-examiner-style critical review of a patent application	Codex MCP
/jurisdiction-format	Compile patent application into jurisdiction-specific filing format (CN / US / EP)	None

(Orchestrator /patent-pipeline chaining all of the above lives under Workflow Orchestrators.)

🧰 Meta, Utilities & Integrations

Cross-cutting infrastructure used by other skills or run on demand.

Skill	Role	Requires
/research-wiki	Persistent research knowledge base — papers / ideas / experiments / claims with typed relationships. Workflow hooks auto-ingest across the research lifecycle	None (pure Python stdlib)
/wiki-enrich	Fill the per-paper TODO sections that ingest_paper leaves as scaffolds (Karpathy LLM-wiki principle). Fetch chain alphaxiv → deepxiv → arXiv → page abstract; idempotent by default, --force to rewrite	Python stdlib + WebFetch
/render-html	Render ARIS Markdown / JSON artifacts into reviewed single-file HTML views for human reading	Python stdlib; Codex MCP for review gate
/overleaf-sync	Two-way sync between local paper directory and Overleaf project via Overleaf Git bridge (Premium) — setup / pull (diff protocol) / push (confirmation gate) / status	Overleaf Premium + macOS Keychain
/feishu-notify	Send notifications to Feishu / Lark — push-only (webhook) or interactive (bidirectional) modes. Off by default	Feishu webhook URL
/interview-cheatsheet	Generate long-form Chinese ML / LLM interview-prep cheat sheets with formulas, code, Q&A, review, and HTML output	Codex MCP, Python

---

How to use this catalog

• Looking for a workflow entry point? Start with Workflow Orchestrators.
• Want to add a skill to an existing workflow? Read the orchestrator's SKILL.md to see which sub-skills it composes.
• Building your own pipeline? Pick the skills from each category and chain them via prompt — no framework lock-in, every skill is a single SKILL.md readable by any LLM agent.

Adding a new skill

1. Create skills/<name>/SKILL.md with name: + description: frontmatter (the description shows up in the LLM's slash-command autocomplete).
2. Per the integration-contract.md §2 contract, if your skill invokes any helper script under tools/, use the canonical resolver chain — do NOT hardcode python3 tools/foo.py.
3. Add a row to the appropriate category table above (or propose a new category in your PR if your skill doesn't fit).
4. The advisory CI lint will catch any hardcoded-path regressions on PR.

See the main README for installation, setup, and end-to-end workflow examples.