ai-research-skills

Fresh-context cross-model review protocol that prevents self-congratulatory score inflation - routes verdict-bearing reviews (idea critique, paper review, claim audits) to a different model or a zero-context fresh thread with a bias guard. Use when running any review/audit step of the pipeline, when the user asks for an unbiased second opinion, cross-model review, or mentions review score inflation.

Orchestrate the full research lifecycle (literature → ideation → novelty check → idea critique → experiments → paper writing → review simulation → rebuttal → thesis) with human decision cards between stages. Use when the user asks to start/resume a research project, run the research pipeline, or asks "what's the next step" in an ongoing research effort. Do not use for a single isolated task (use the stage-specific skill directly).

Check whether a proposed research novelty (given a problem statement and claimed idea/novelty) overlaps with existing published work. TRIGGER when the user asks to verify research novelty, check if an idea is new, compare a proposed contribution to prior art, do a literature/prior-art search for a specific claim, or assess whether a paper idea has been scooped. DO NOT TRIGGER for general literature reviews unrelated to a specific novelty claim, or for writing related-work sections of an already-validated idea.

Select and run the correct hypothesis test based on data properties. Covers parametric/non-parametric tests, effect sizes, and multiple comparison correction.

Use when main results pass result-to-claim (claim_supported=yes or partial) and ablation studies are needed for paper submission. Do not use for general experiment design (use experiment-design).

Analyze ML experiment results, compute statistics, generate comparison tables and insights. Use when user says "analyze results" or needs to interpret experimental data. Do not use for run-vs-run alignment and tracking (use compare).

Autonomously improve a generated paper via GPT-5.6-Sol xhigh review → implement fixes → recompile, for 2 rounds. Use when user says "改论文", "improve paper", "论文润色循环", "auto improve", or wants to iteratively polish a generated paper.

Autonomous multi-round research review loop. Repeatedly reviews via external reviewer backend (Codex or manual), implements fixes, and re-reviews until positive assessment or max rounds reached. Use when user says "auto review loop", "review until it passes", or wants autonomous iterative improvement.

SSH job queue for multi-seed/multi-config ML experiments with OOM-aware retry, stale-screen cleanup, and wave-transition race prevention. Use when user says "batch experiments", "队列实验", "run grid", "multi-seed sweep", "auto-chain experiments", or when /run-experiment is insufficient for 10+ jobs that need orchestration.

Draft a structured grant proposal from research ideas and literature. Supports KAKENHI (Japan), NSF (US), NSFC (China, including 面上/青年/优青/杰青/海外优青/重点), ERC (EU), DFG (Germany), SNSF (Switzerland), ARC (Australia), NWO (Netherlands), and generic formats. Use when user says "write grant", "grant proposal", "申請書", "write KAKENHI", "科研費", "基金申请", "写基金", "NSF proposal", or wants to turn research ideas into a funding application.

Adversarial attack-defense exercise on a near-final paper - a fresh reviewer writes the single strongest 200-word rejection memo, a second fresh reviewer defends point-by-point, and an adjudicator classifies each attack point as answered / partially answered / still unresolved. Use once before submission after standard reviews have stabilized, or when the user says kill argument, hostile review, or asks what would make an area chair reject this paper.

Unattended overnight research iteration - runs bounded experiment/tuning loops with objective machine-checkable stop conditions, safety rules, per-iteration logging, and a morning report, while explicitly excluding verdict-bearing decisions from automation. Use when the user asks to run experiments overnight, keep iterating unattended, sweep parameters while away, or set up an autonomous research loop.

Zero-context verification that every number, comparison, and scope claim in the paper matches raw result files. Uses a fresh cross-model reviewer with NO prior context to prevent confirmation bias. Use when user says "审查论文数据", "check paper claims", "verify numbers", "论文数字核对", or before submission to ensure paper-to-evidence fidelity. Do not use for code/data/formula verification (use paper-verification).

Generate publication-quality AI illustrations for academic papers via pluggable image backends (Gemini / OpenAI gpt-image / any OpenAI-compatible endpoint, with mermaid fallback). Creates architecture diagrams, method illustrations with Claude-supervised iterative refinement loop. Use when user says "生成图表", "画架构图", "AI绘图", "paper illustration", "generate diagram", or needs visual figures for papers.

Build a one-off HTML single-page academic poster with measurement-driven hard gates, real paper figures, a two-hue design-token system, and print-ready PDF via headless Chromium. Use only for "HTML 单页海报" or "one-off HTML poster" requests. Do not use for general poster requests (use paper2poster).

Generate conference presentation slides (beamer LaTeX → PDF + editable PPTX) from a compiled paper, with speaker notes and full talk script. Use when user says "做PPT", "做幻灯片", "make slides", "conference talk", "presentation slides", "生成slides", "写演讲稿", or wants beamer slides for a conference talk.

Render an ARIS Markdown / JSON artifact (IDEA_REPORT, AUTO_REVIEW, KILL_ARGUMENT, PAPER_PLAN, research-wiki state, etc.) into a single-file HTML view designed for human reading. Use when the user says "渲染 HTML", "出一份 HTML 报告", "render html", "make this readable", "export to html", or wants a polished web-rendered view of a Markdown artifact.

Use when the user wants to prepare code for open-source release, create reproducible research artifacts, or structure a repository for publication. Triggers on phrases like "publish code", "open source release", "reproducibility", "research repository", "code release", or "prepare for publication".

Persistent research knowledge base that accumulates papers, ideas, experiments, claims, and their relationships across the entire research lifecycle. Inspired by Karpathy's LLM Wiki pattern. Use when user says "知识库", "research wiki", "add paper", "wiki query", "查知识库", or wants to build/query a persistent field map.

Use when experiments complete to judge what claims the results support, what they don't, and what evidence is still missing. Codex MCP evaluates results against intended claims and routes to next action (pivot, supplement, or confirm). Use after experiments finish — before writing the paper or running ablations.

Autonomous research review loop using any OpenAI-compatible LLM API. Configure via llm-chat MCP server or environment variables. Trigger with "auto review loop llm" or "llm review".

Zero-context verification that every cited bibliography entry is real, correctly attributed, and actually supports the sentence citing it - catching hallucinated references, wrong metadata, and wrong-context citations with per-entry KEEP/FIX/REPLACE/REMOVE verdicts. Use before any submission (paper or thesis) or when the user asks to check citations, verify references, or audit the bibliography.

Use when the user wants to analyze dataset bias, create stratified samples, evaluate fairness, or plan dataset collection. Triggers on phrases like "dataset bias", "stratified sample", "class imbalance", "data distribution", "fairness analysis", or "ethical review".

Hypothesis-driven experiment loop with fair-comparison discipline - reproduce baselines, run minimal validation, iterate via hypothesis tree, and build the ablation/evidence base for a paper. Use when the user asks to run experiments, validate an idea, reproduce a baseline, design ablations, or improve a method iteratively. Do not use for pure idea generation (idea-mining) or paper drafting (paper-writing).

Monitor running experiments, check progress, collect results. Use when user says "check results", "is it done", "monitor", or wants experiment output.

Use when the user wants to verify paper claims against code or data, audit numerical accuracy, check formula-code alignment, or validate citation accuracy. Triggers on phrases like "verify claims", "check numbers", "do the numbers match", "formula vs code", "audit the paper", or "cross-check results". Do not use for paper-only claim wording and citation audit (use paper-claim-audit).

Improve academic paper writing quality for ML/CV/NLP-style papers with clear section structure, paragraph flow, and reviewer-facing presentation. Use when drafting or revising Abstract, Introduction, Related Work, Method, Experiments, or Conclusion; polishing figures/tables; checking claim-support alignment; or performing self-review before submission.

Render a pre-extracted paper's structured 9-section spec (`paper_spec.md`) into a single-page HTML academic poster, fit the layout to the page via an iterative measured-fill loop, and export it to print-ready PDF + PNG thumbnail. Requires the upstream `paper2assets` skill to have produced the input `<outdir>/` package (`manifest.json` at the root + an `assets/` folder holding `meta/paper_spec.md`, `meta/text.txt`, `meta/figures.json`, `meta/metadata.json`, `figures/*.png`, `logos/`, `qr/`) first. Use when the user wants an HTML poster, PDF/PNG export, or PPTX from a paper they already have extracted assets for — e.g., "render the poster", "make the poster from this spec", "export poster to PDF", "paper2poster". The three skills paper2assets → paper2poster → html2pptx run in sequence, each invokable on its own.

Use this Skill to write structured academic peer-review rebuttals: point-by-point responses, tone calibration, LaTeX templates, and cover letters for re-submission. Also read references/rebuttal-strategy.md for concern triage (misunderstanding / fixable / fundamental), rebuttal-experiment scheduling, and score-raising strategy.

Deploy and run ML experiments on local, remote, Vast.ai, or Modal serverless GPU. Use when user says "run experiment", "deploy to server", "跑实验", or needs to launch training jobs.

Convert one or more published/submitted conference papers into a master's thesis - restructuring from paper format to thesis chapters, expanding background and related work, unifying notation, and complying with the university LaTeX template and Chinese academic writing norms. Use when the user asks to turn papers into a thesis, write thesis chapters, or prepare 学位论文.

Lightweight pattern-driven research ideation with no script dependencies - generates 5-10 candidate ideas from a direction or gap matrix using proven ideation patterns, each with mechanism, falsification plan, and minimal validation experiment. Use for quick brainstorming of research ideas or when the scripted idea-spark skill is unavailable or not yet set up; prefer idea-spark for a single deeply-grounded reviewer-defensible idea card. Do not use for judging an existing idea (use idea-critic) or checking novelty (use scoop-check).

Systematic literature survey plus research-gap mining. Builds a structured literature table and a method-problem-dataset gap matrix, surfacing contradictions and unexplored cells that seed research ideas. Use when the user asks for a literature review, survey of a research area, "what has been done", or to find research gaps/opportunities. Do not use for checking one specific novelty claim (use scoop-check).

Same-epoch comparison of training runs across wandb, neptune, tensorboard, or mlflow. Aligns runs at the student's current step (never current-vs-final-of-baseline) and separates proxy metrics from downstream targets. Use when the user asks to compare runs, check if a run is improving, track lag against a baseline, rank experiments, or evaluate run-vs-run performance. Do not use for statistical interpretation of completed result tables (use analyze-results).

Use when the user wants to design experiments, plan ablation studies, structure baselines, or create incremental evaluation strategies. Triggers on phrases like "design ablation", "plan experiment", "what experiments should I run", "baseline comparison", or "experiment matrix". Do not use for ablation-planner requests after main results pass result-to-claim (use ablation-planner).

Judge whether a research idea contains genuine insight (vs. combinatorial stitching), passes domain-intuition sanity checks, and would survive peer review. Three-layer critique - insight tests, domain pitfall checks, multi-persona reviewer simulation - producing a structured critique card with an accept/pivot/abandon verdict. Use when the user asks to evaluate, critique, or stress-test a research idea before committing to it. Do not use for finished paper review (use reviewer-simulation) or novelty search (use scoop-check).

Generate ONE reviewer-defensible, implementable research idea with a concrete method and falsification plan from a stated research direction. Use when the user asks for a research idea, novelty analysis, bottleneck diagnosis, or paper-shape suggestion. Skip code review, debugging, and unconstrained brainstorming without research context.

Search papers across arXiv, DBLP, OpenAlex, OpenReview, Semantic Scholar, and Crossref for a given query and year range, using ./scripts/search_papers.py. Use when the user asks to find papers, related work, prior art, or recent publications on a specific topic, especially when they mention a date range or specific venues like NeurIPS, ICLR, or ICML.

End-to-end conference talk pipeline: paper → slide outline → Beamer + PPTX → per-page polish → assurance checks (claim / citation / anonymity) → final export and report. Default-good for academic conference talks (NeurIPS / ICML / ICLR / VALSE / 投稿 talks). Trigger phrases: "做 talk", "做 PPT 全流程", "talk pipeline", "end-to-end slides", "做演讲", "conference talk full workflow". Use when the user wants the complete talk artifact, not just a slide deck.

Extract a research paper PDF into a structured set of poster-agnostic assets reusable by any downstream renderer (paper2poster, paper2blog, paper2audio, paper2video). Produces a `<outdir>/` containing the paper's full text (assets/meta/text.txt), per-figure captions (assets/meta/captions.json), cleaned figure rasters (assets/figures/*.png + assets/meta/figures.json manifest), paper metadata (assets/meta/metadata.json: title / authors / institutes / venue / paper_url / code_url), institute logos (assets/logos/*), URL QR codes (assets/qr/*), and a 9-section structured paper summary (assets/meta/paper_spec.md). The bundle follows the Output Contract layout (deliverables at top, everything else under `assets/`). Use when the user wants to extract paper content into reusable assets, OR as the mandatory upstream stage of any paper-rendering pipeline — e.g., 'extract this paper', 'build paper assets', 'get the figures and spec from this PDF', 'paper2assets'.