open-paper-skills

Maintain an OKF paper wiki under `paper-wiki/` or another user-named wiki root, including optional Chinese mirror notes under `papers_zh/`. Use when asked to add arXiv or research-paper URLs, capture non-paper sources such as blog posts, docs, or talks, update or translate paper notes, maintain paper/topic/concept/source indexes, automatically create or update topic summary pages for important new themes, maintain concept entity pages for methods, datasets, benchmarks, metrics, terms, or tools, link papers to research projects, normalize paper metadata, track reading status, compare papers, generate required `viz.html` visualizations, or validate the paper wiki.

Audit, position, and communicate research value without outrunning the evidence. Use when users need audience-specific buy-in from advisors, collaborators, reviewers, talk audiences, or the research community; want to pitch or promote an idea, paper, result, codebase, or demo; plan research PR; or need to detect and repair overclaiming. Do not use for generic proofreading, summarization, or formatting.

Generate or lint a compact top-level FILETREE.md navigation map. Use when repository structure changes or FILETREE.md needs creation, regeneration, or validation.

Map scattered ideas into a human-approved DAG of verifiable states, then drive fresh agents through it edge by edge. Use when the user has a destination or idea fragments but lacks the domain knowledge to identify or sequence the route; when revising a map bundle; when compiling approved edges into launch prompts; or when recording execution evidence, deviations, and human verification.

Prompt-contract authoring for frontier reasoning and coding agents. Use when the user wants a new implementation prompt, a phased agent task, a copy-ready prompt, or an evidence-based upgrade of an existing prompt that should work across capable model providers.

Two-stage academic rebuttal assistant. Given the paper abstract, reviewer comments, scores, confidences, and venue constraints, first diagnose the underlying concerns, assess whether rebuttal is worth the limited author time, and produce a prioritized experiment-and-analysis plan. After the author provides completed results, produce an evidence-grounded rebuttal, revision plan, or resubmission plan. Never fabricate results, reviewer intent, citations, or manuscript locations.

Use when selecting, auditing, or justifying experimental baselines for a research idea, paper draft, task, benchmark, or experiment plan. Finds classic methods, recent SOTA, direct competitors, strong simple baselines, ablation baselines, resource-matched comparisons, and reviewer-expected missing baselines, but only admits baselines with non-empty, accessible GitHub repositories that appear reproducible. Records strong papers without usable open-source code as excluded candidates rather than selected baselines.

A protocol framework for long-horizon autonomous tasks. Targets three empirically-observed failure modes — cognitive loops, stalling, runtime fragility — by prescribing state management, stall detection, and watchdog mechanisms. Validated on multiple task types including paper writing (4 ICLR-format surveys, in-framework self-rating 8.0-8.6/10).

Discover research/academic agent-skills from skills.sh (and user-supplied candidates), filter them through strict hard gates plus a strict academic-relevance gate, score survivors with a rubric, and hand back a ranked report with reasons. Use when the user wants to find, discover, or scout new academic/research skills to add to the hub, evaluate a skill they saw on social media, or refresh the discovery ledger.

Scaffold human cognition with a project-local, git-trackable four-quadrant cache. Use when (1) initializing, reading, updating, or migrating the cache; (2) completing a task beyond the human's confirmed understanding, where cognition should guide agent autonomy and decision escalation; or (3) explaining a verified artifact at the human's level and recording evidence-backed cognitive growth.

Organize a repository, folder, or knowledge corpus into generic Open Knowledge Format (OKF) bundles. Use when you are asked to convert or normalize repo files and folders to OKF, add OKF index/log/concept frontmatter, choose bundle boundaries, migrate idea/project/doc folders into OKF-style structure, or validate generic OKF conformance.

Research operating-system guidance for ML/AI research. Use this skill when you need to help choose research problems, design experiment loops, improve research taste, debug model training or eval workflows, inspect failures, keep research logs, build reading plans, or turn how_to_be_good_at_research.md principles into concrete plans, checklists, habits, or advice.

Use when starting or reviewing a complex implementation where the user wants an agent to interrogate the plan against docs/source evidence, surface unknown unknowns, and avoid rushing into build mode. Combines docs-grounded grilling with a map-vs-territory unknowns pass.

Generate publication-quality academic paper diagrams (.drawio) and statistical plots (.png via Python/matplotlib) using a PaperBanana-inspired multi-agent pipeline. Reads methodology sections, figure captions, and raw data to produce structured, aesthetically refined visualizations with reference-driven design.

Use when the user asks for a rich explanation of a research paper, blog post, article, documentation, or any complex concept. Produces HTML output. Use this skill whenever the user wants to deeply understand or explain a paper, blog, or hard idea — e.g. "explain this paper", "help me understand this blog post", "break down this concept" — even if they don't explicitly ask for HTML.

Andrej Karpathy's field notes (CLAUDE.md, v260626) on getting a language model to write code you will NOT have to rewrite — ten earned rules covering read-before-write, think-before-code, simplicity, minimal/surgical diffs, verification, goal-driven execution, debugging, dependency hygiene, communication, and the common failure modes. Read and apply these BEFORE starting ANY coding task: writing new code, editing or refactoring existing code, fixing a bug, adding a feature, implementing from a spec, reviewing a diff, or scaffolding a project — even when the user does not explicitly ask for "best practices" or "clean code." Whenever the task is to produce or change source code, consult this skill first so the rules shape how you read, plan, write, verify, and report.

Interactively draft a well-structured `task.md` brief for Claude Code by walking the user through Objective, Acceptance Criteria, Relevant Files, Constraints, Background, Known Context, and Questions based on Opus 4.7 best practices. Use this skill whenever the user wants to prepare a task, brief, spec, prompt, plan, or instructions for Claude Code to execute in a fresh session — even if they don't explicitly ask for a "task file" or "task.md". Trigger on phrases like "help me write a task", "draft instructions for Claude Code", "prepare a brief", "plan what to ask Claude to do", "turn this idea into a task.md", "write a spec for Claude", or "front-load context for a new Claude Code session".

Create and manage Python project environments with the uv package manager. Use whenever the user wants to set up a new Python project, create a virtual environment, initialize a pyproject.toml, install/add Python packages, lock or sync dependencies, or pin a Python version — even if they don't explicitly mention "uv". Prefer this skill over raw pip/venv/conda for any Python environment work in a project directory.

Look up any arxiv paper on alphaxiv.org to get a structured AI-generated overview. This is faster and more reliable than trying to read a raw PDF.

Generate draw.io diagrams as .drawio files, optionally export to PNG/SVG/PDF with embedded XML

Write publication-ready ML/AI papers for NeurIPS, ICML, ICLR, ACL, AAAI, COLM. Use when drafting papers from research repos, structuring arguments, verifying citations, or preparing camera-ready submissions. Includes LaTeX templates, reviewer guidelines, and citation verification workflows.