Claude Skills
The open format is called Agent Skills and works in Claude Code, Codex, Cursor and other agents — most people know it as Claude Skills.
Every Agent Skill we could find on GitHub, deduplicated by content. 80 149 files from 1 774 authors, of which 62 489 are unique — the rest is the same skill repackaged into someone else's repository. For each one: what it weighs in tokens, whether it ships runnable scripts, and which MCP servers it needs.
> Structure prediction using Chai-1, a foundation model for molecular structure. (2) Validating designed binders, (3) Predicting protein-ligand complexes, (4) Using the Chai API for high-throughput prediction, (5) Need an alternative to AlphaFold2. For QC thresholds, use protein-qc. For AlphaFold2 prediction, use alphafold. For ESM-based analysis, use esm.
> Ligand-aware protein sequence design using LigandMPNN. (2) Enzyme active site design, (3) Ligand binding pocket optimization, (4) Metal coordination site design, (5) Cofactor binding proteins. For standard protein design, use proteinmpnn. For solubility optimization, use solublempnn.
> Binder design ranking using ipSAE (interprotein Score from Aligned Errors). (2) Filtering BindCraft or RFdiffusion outputs, (3) Comparing AF2/AF3/Boltz predictions, (4) Predicting binding success rates, (5) Need better ranking than ipTM or iPAE. For structure prediction, use chai or alphafold. For QC thresholds, use protein-qc.
> End-to-end guidance for protein design pipelines. (2) Need step-by-step workflow guidance, (3) Understanding the full design pipeline, (4) Planning compute resources and timelines, (5) Integrating multiple design tools. For tool selection, use binder-design. For QC thresholds, use protein-qc.
> (1) Designing sequences for RFdiffusion backbones, (2) Redesigning existing protein sequences, (3) Fixing specific residues while designing others, (4) Optimizing sequences for expression or stability, (5) Multi-state or negative design. For backbone generation, use rfdiffusion or bindcraft. For ligand-aware design, use ligandmpnn. For solubility optimization, use solublempnn.
> Generate protein backbones using RFdiffusion, a diffusion-based generative (1) Designing binder scaffolds for a target protein, (2) Generating novel protein backbones from scratch, (3) Scaffolding functional motifs into new proteins, (4) Specifying hotspot residues for interface design, (5) Creating symmetric oligomers. For sequence design after backbone generation, use proteinmpnn. For structure validation, use alphafold or chai. For QC thresholds, use protein-qc.
> Structure prediction with Protenix, an open AlphaFold3 reproduction. Use this (2) Wanting an open alternative to AF3 alongside Boltz and Chai, (3) Validating designed binder-target complexes. For QC thresholds, use protein-qc. For ipSAE ranking, use ipsae.
> Quality control metrics and filtering thresholds for protein design. (2) Setting filtering thresholds for pLDDT, ipTM, PAE, (3) Checking sequence liabilities (cysteines, deamidation, polybasic clusters), (4) Creating multi-stage filtering pipelines, (5) Computing PyRosetta interface metrics (dG, SC, dSASA), (6) Checking biophysical properties (instability, GRAVY, pI), (7) Ranking designs with composite scoring. This skill provides research-backed thresholds from binder design competitions and published benchmarks.
> (1) User is new and hasn't run any tools yet, (3) Modal authentication errors occur, (4) User asks how to get started or set up the environment, (5) biomodals directory is missing or tools aren't working.
> Access UniProt for protein sequence and annotation retrieval. (2) Finding functional annotations, (3) Getting domain boundaries, (4) Finding homologs and variants, (5) Cross-referencing to PDB structures. For structure retrieval, use pdb. For sequence design, use proteinmpnn.
> Solubility-optimized protein sequence design using SolubleMPNN. (2) Optimizing solubility of designed proteins, (3) Reducing aggregation propensity, (4) Need high-yield expression, (5) Avoiding inclusion body formation. For standard design, use proteinmpnn. For ligand-aware design, use ligandmpnn.
Transform real Chinese architecture and place-based cultural photos into art-directed editorial posters, integrated multi-photo scenes, and optional source comparisons. Use for historic buildings, vernacular streets, culturally rooted shops, interiors, food, travel covers, and place montages; do not use for generic advertising, routine retouching, or invented places.
>- 质检 Agent：检查 vibe-coding-kit 其他 Skill 的产出物是否符合治理规范。 当用户说"检查一下产出"、"帮我看看合不合规"、"质检"、"验证 PRD"、"检查项目说明书"， 或在任何 Skill 完成产出后，使用此 Skill 做合规检查。 它读取项目根目录的 harness.json，逐条校验产出文件，报告违规项和修复建议。 这是 vibe-coding-kit 套件的治理层（harness），确保所有产出物格式统一、质量达标。 它不止检查格式，还会做跨 Skill 一致性校验——比如 PRD 里的项目名和项目说明书里的项目名是不是一致。
>- Vibe Coding（用 AI 写代码）开发过程中的避坑与自救——真正翻车大多发生在"开工后"。 当用户说"AI 越改越乱"、"demo 改坏了退不回去"、"AI 老忘记之前说的"、"功能越加越多很乱"、 "AI 说做好了但其实没用"，或者正在用 AI 持续开发一个项目时，主动使用此 Skill。 涵盖：管理 AI 对话（防忘事）、守住范围（防膨胀）、保住能用的版本（防丢失）、让 AI 证明给你看（防轻信）， 以及哪些事必须找真人工程师的「红线」，和把一切串起来的「项目说明书」。 这是 vibe-coding-kit 套件里贯穿整个开发过程的 Skill，建议从第一行代码起就配合使用。
>- 把验证过的 Vibe Coding demo，做成能长期运行、给别人用的正式系统。 当用户说"想上线"、"怎么部署"、"这个 demo 想做成正式的"、"要注意安全吗"、"怎么测试我的项目"， 或者准备把代码开源/公开发布时，使用此 Skill。涵盖开发规范、安全基线、部署、手动验收、文档要求。 这是 vibe-coding-kit 套件里负责"从 demo 到上线"的 Skill。 即使用户没明说"上线"二字，只要 ta 准备把一个能跑的东西交给别人用、或放到服务器/公网上，就应主动用本 Skill。
>- 一个"访谈式"的需求 Agent：通过一问一答 + 在岔路口给建议，把一个说不清需求的人，带到一份 AI 能直接照着开工的 PRD。 当用户说"我想做个 XX 但不知道怎么跟 AI 说"、"帮我理需求"、"帮我写个 PRD"、"我有个想法，帮我想清楚再开工"， 或在 Claude Code / Codex 里准备开一个新项目、却还没把要做什么定清楚时，使用此 Skill。 它不是让用户填表格，而是像产品顾问一样访谈：一次只问一件事，遇到用户判断不了的选择（尤其"产品形态"和"技术栈"） 会给出 2-3 个大白话选项 + 推荐一个最简可行的默认 + 一句话讲清代价，等用户拍板。 访谈结束后，直接把 prd.md 和 项目说明书.md 写进项目目录，并提议生成可点击的 HTML 原型。 这是 vibe-coding-kit 套件里"把需求从人嘴里问出来"的交互入口。
>- Vibe Coding（用 AI 写代码）项目的第一步：把模糊的想法，变成 AI 能精准落地、不会跑偏的需求。 当用户说"我想做一个 XX"、"帮我用 AI 做个 demo"、"我有个想法想验证"、"帮我跟 AI 对齐需求"， 或者描述了一个还很模糊的软件/工具/系统点子时，使用此 Skill。也用于先判断"这件事到底要不要写代码"。 这是 vibe-coding-kit 套件的入口——只想跑 demo 验证想法的人，通常只需要这一个 Skill。 它不止教你说清"想要什么"，还教你用"数据旅程 + 3×4 提问"把需求补全到考虑周全，避免只描述了"一切顺利"的happy path。 即使用户没明说"规划"二字，只要 ta 准备让 AI 开始写代码、却还没把需求说清楚，就应主动用本 Skill 先做需求对齐。
>- 教非技术背景的人看懂技术架构、做好技术选型——把"只会照搬话术的小白"练成"能看穿方案好坏的人"。 当用户问"该用什么技术/框架"、"帮我选技术栈"、"AI 给的方案我看不懂"、"这几个方案怎么选"， 或者需要把 demo 做成更正经的东西、要理解 AI 提出的架构时，使用此 Skill。 这是 vibe-coding-kit 套件里负责"成长"的核心 Skill：用 6 个大白话维度拷问任何方案，再配五步决策法做选择。 即使用户没明说"架构"二字，只要 ta 面对多个技术方案不知道怎么选、或看不懂 AI 的技术建议，就应主动用本 Skill。
Update agent context files for this project. Use when the user invokes /update-agent-context, asks to "update agent docs", "update CLAUDE.md", "update AGENTS.md", "add to agent_docs", "record a correction", or mentions updating context for a specific topic.
Set up Claude Code for this project. Use when the user invokes /setup-claude, asks to "set up claude", or "create the skills symlink".
Run linting, formatting, and type checking on changed files. Use when asked to "lint", "fix linting", "run linter", "check formatting", or "check types".
Lightweight guidance for using the Upstash Workflow SDK to define, trigger, and manage workflows. Use this Skill whenever a user wants to create workflow endpoints, run steps, or interact with the Upstash Workflow client.
终极口播视频 skill：中文口播稿 + 成品配音 → CPU 字级时间戳 → SHOTBOOK 层矩阵分镜 → Remotion 电影感成片（横屏默认/竖屏）。当用户要"做口播视频"、"解说/科普视频"、"把文案变成视频"、"给配音配画面动效"时使用。TTS 合成与数字人生成技术不在本 skill 内（配音和人物素材是输入）。含统一视觉语言（Apple 范式）、78 张动效配方卡、镜头三面分层工作单、七层镜头反PPT系统（CameraRig/视差/让位/环境）、六式运动承接转场（每式一卡）、长镜头世界画布、anime.js+three.js 桥、自动静止检测 + 独立 subagent 评估循环。
人前で話す登壇・講義スライドのストーリーの組み方。つかみ、中扉、段階的な開示、見出しの文体、締め方、尺の見積り。「登壇スライドを作って」「LTの資料」「講義資料の構成を考えて」と言われたら、書き始める前に読む。
登壇スライドに載せる図・構成図・挿絵の作り方。情報量の絞り方、SVGの描き方、文字サイズの下限、挿絵の置き方、書き出し後の検査。「図を作って」「構成図を描いて」「図が見づらい」「挿絵を入れて」と言われたら読む。
黒地のMarpテーマ（minorun-dark）でスライドを組むときのデザインバランスと検査。余白の測り方、縦のバランス、配色の決め方、表・コード・補足ボックスの確定デザイン、Marp固有の罠、書き出し後の検査手順。「スライドのバランスを整えて」「色を変えて」「余白が変」「黒地のテーマで作って」と言われたら読む。
>- wikilinks, orphan pages, empty stubs, frontmatter that violates the schema, near-duplicate pages, and stale index entries. Use this skill whenever the user asks to lint, clean, audit or check the health of their vault, says pages feel messy or links are broken, after a large bulk import, or when a scheduled maintenance run fires. Do NOT use for ingesting new sources, for answering questions from the vault, or for rewriting page content that is structurally fine but reads badly.
>- read the source, split it into concepts and entities, write or update pages, connect them to existing pages, and record the run in the log. Use this skill whenever the user drops a file into raw/, pastes an article, transcript or PDF and asks to add it to the vault, says "ingest this", "add this to my second brain", "process raw", or asks to catch up on unprocessed sources, even if they do not name the ingest command. Do NOT use for answering questions from an existing vault, for linting or repairing pages, or for editing notes the user wrote by hand.
>- claims, contradictions between sources of different dates, and what evidence moved them. Use this skill when the user asks what they have changed their mind about, how their thinking on a topic developed, what they used to believe, or wants a quarterly look back at their own positions. Do NOT use for finding contradictions between sources with no time dimension, or for a current-state answer.
>- worth ingesting, process oldest first, checkpoint after every batch, and keep cost visible. Use this skill whenever the user wants to import years of bookmarks, an export from another notes app, a downloads folder, a chat history dump, or says "backfill", "bulk import", "process my archive". Do NOT use for ingesting one or a few new sources, which is second-brain-ingest.
>- Merge two pages that describe the same thing, preserving every distinct claim, redirecting inbound links, and recording the merge. Use this skill when the user approves a duplicate merge, says "these are the same page", "merge X into Y", or asks to clean up near-duplicates that a lint run proposed. Do NOT use to merge pages on your own judgement, to delete a page outright, or to split one page into several.
>- Record a dated snapshot of the vault's health numbers and compare against previous snapshots so trends are visible. Use this skill for a monthly or quarterly check, when the user asks whether the vault is improving, or asks to track its numbers over time. Do NOT use for a one-off graph analysis, which is second-brain-graph, or for structural repairs.
>- degree, components, hubs, bridges and clusters, with what each number means for retrieval. Use this skill when the user asks about the health or structure of their graph, wants metrics, asks which pages are most connected or most isolated, or asks whether the vault is fragmented. Do NOT use for fixing what it finds, which is second-brain-lint, or for answering a content question.
>- Triage an exported chat history into what to ingest, what to archive and what to delete, with a privacy pass first. Use this skill when the user has exported their Claude or ChatGPT history, asks to import old conversations, or asks what in their chat archive is worth keeping. Do NOT use for ingesting a single pasted conversation, or for any material the user has not explicitly asked to bring into the vault.
>- material, and report what would leak through links before anything is exported. Use this skill when the user wants to publish notes as a site or digital garden, asks what is safe to make public, or is setting up an export. Do NOT use for exporting a single document, or for running the publishing tool itself.
>- Generate questions from the user's own concept pages so they can find out what they saved but never learned, and compare their answers against what the pages say. Use this skill when the user wants to test themselves, revise, make flashcards from their notes, or asks what they have forgotten. Do NOT use for generic quizzing on a topic the vault does not cover, or for answering a question from the vault.
>- pages, name the gaps, and write findings with citations, disagreements and limits. Use this skill when the user asks for a research report, a deep write-up on a topic, or wants to understand a subject thoroughly using what they have collected. Do NOT use for a quick answer from the vault, for drafting an article in the user's voice, or for summarising a single source.
>- project-scoped CLAUDE.md with one goal, and links to the concept pages it depends on. Use this skill when the user starts something new they will work on over time, says "create a project", "set up a folder for X", or asks to define the goal and structure of ongoing work. Do NOT use for one-off tasks, for wiki pages about a topic, or for scoping an existing project down to work in it.
>- people's private information, confidential work, and anything the user would not want synced or backed up. Use this skill when the user asks about privacy, is about to share or sync the vault, has just imported chat history or meeting notes, or asks what is sensitive in their notes. Do NOT use to delete anything on your own, or as a substitute for the publishing check.
>- topics are growing, what contradicts what, which questions are still open, and what is worth reading next. Use this skill whenever the user asks for a weekly or monthly review, asks what changed in their vault, what they have been learning about lately, or what they should look at next, and when a scheduled review task fires. Do NOT use for structural linting, for ingesting sources, or for answering a specific factual question.
>- Answer a question from a second-brain vault by reading its own pages, citing the pages used, and saying plainly when the vault does not contain the answer. Use this skill whenever the user asks what they know about a topic, what they have read on something, what connects two ideas, what they have changed their mind about, or asks any question that should be answered from their own notes rather than general knowledge. Do NOT use for ingesting new material, for general questions the vault has nothing to do with, or when the user explicitly wants an answer from the open web.
>- name as an alias, and log it. Use this skill whenever a page needs a different canonical title, the user says "rename X to Y", asks to fix an inconsistent title, or after a lint run flags naming problems. Do NOT use for merging two pages, for moving a page between folders without a name change, or for renaming files outside the wiki.
>- fix mistranscribed technical terms, and split long recordings by topic. Use this skill when the user drops an auto-generated transcript, subtitle file, podcast or lecture transcript, meeting recording text, or asks to process a video or audio source. Do NOT use for text that is already prose, for summarising a transcript, or for the ingestion itself.
>- own sources, and surface where their material disagrees. Use this skill when the user wants to write an article, post, essay or newsletter based on what they have collected, or asks what they could write about a topic. Do NOT use for research reports, for writing with no grounding in the vault, or for editing text the user wrote themselves unless asked.
Create dark-themed, animated technical diagrams as self-contained HTML+SVG files — flowcharts whose connectors visibly flow, and architecture diagrams where requests travel as light dots through the system (Diagrid/Temporal landing-page style). Use this skill whenever the user asks for a flowchart, workflow, pipeline, process diagram, state machine, system architecture, infrastructure, cloud, microservices, or network topology diagram — and especially when they mention "animated", "flowing", "dynamic", "alive", "GIF-like", or want a diagram for a landing page, README, docs, or product demo. Also use it to convert Mermaid source (a mermaid code block or .mmd file) into an animated diagram — "animate this mermaid", "make this flowchart move". Prefer this over static diagram output whenever the diagram represents anything that moves: requests, events, data, jobs, messages, or control flow.
Read editor.md, editor.json, and analyst.json. Make data-driven creative visual decisions for every section — teaser, charts, images, videos, audio, maps, and interactive demos when they fit the data. Generate selected assets. No HTML. Outputs designer.json with des_xx IDs.
Name the piece — re-write the masthead (headline + standfirst + kicker), every section title, and every figure/photo/table caption to a research-driven titling standard, killing the AI-tell patterns (the 'flat statement. flat counter-statement.' two-beat above all) a competent default falls into. Reads editor.md/json + analyst.json + the resolved topic_profile; writes copywriter.json — STRINGS ONLY (masthead{headline,standfirst,kicker}, items{edt_xx:{title}, des_xx:{caption}}), each backed by a real ana_*. Names, never edits: it touches no finding, no number, no data-* id, no layout — so the Verify layer is untouched and the Programmer renders the masthead + figcaptions from copywriter.json verbatim. Runs at Stage 3.5, after the Editor, before the Designer.
Design the MANDATORY cinematic scroll experience (every blog ships one): a full-bleed, scroll-driven background — primarily from the Scout's verified real imagery — plus the global motion choreography, so the narrative unfolds like a film as the reader scrolls. Owns the BACKGROUND layer + page-level motion (not the Designer's per-section visuals, not the Interaction centerpiece — it stages them inside the scroll). Outputs cinematographer.json (cin_xx scenes). There is no off path: a topic with no sourceable real imagery falls back to a generative-atmosphere or data-driven-spine background, never to a bare column.
Review a finished Data2Story blog against the 5 quality rubric dimensions (visual_design, narrative_pacing, data_method_transparency, claim_data_alignment, insight_value), score each 1-7 with on-page evidence, and emit critic.json with pass/fail + targeted, surgical send-back instructions. Verifies every load-bearing claim/asset against its traceability chain before scoring; applies the caveat-survival, honest-accuracy, and third-party-attribution caps. Does NOT rewrite content — scores and sends back. Use at Stage 6.5 after the Auditor and after verify.py has produced verifier.json; re-runs each revision round. Triggers: a built index.html plus verifier.json exist, or you need to judge whether the article is actually good.
Audit a generated Data2Story blog for build correctness across ALL modalities by ACTUALLY RENDERING it in a real headless browser (when available) — catching blank/0-width charts, broken/oversized media, desktop+mobile overflow, and dead interactions that source inspection cannot see — then having vision agents review the screenshots, playtesting every interactive, and walking the flagship capability contract. Plus VIEW every image for a wrong/AI-faked subject. Fixes layout in place; sends content-correctness problems back to the owning role. Use at Stage 6 after the Programmer assembles index.html and before the Critic; re-runs each revision round. Triggers: a built index.html exists, or you need to confirm charts/media/interactions actually render.
Exhaustively profile a dataset and list ALL possible analyses — distributions, correlations, rankings, trends, group comparisons, anomalies. Reads detective.json for context. Runs after the Detective/Scout, before the Editor. Outputs analyst.json with ana_xx IDs and chart-ready data_tables.
Use to turn a dataset into a verifiable multimedia blog (a data story / data-driven article / interactive dashboard from a dataset). Orchestrator for the Data Journalist Agent (Data2Story): a 7-team newsroom (14 agents) running detective → scout → analyst → imagineer → editor → copywriter → designer → interaction → hero → cinematographer → programmer → auditor → critic → inspector in sequence. Trigger when the user hands over a dataset (CSV/JSON/folder/path) and wants a published story, blog post, or interactive report built from it. Creates a versioned project folder per run.
Generate text embeddings via OpenRouter using Qwen3-Embedding-8B.
Generate atmospheric sound-design / SFX (NOT speech, NOT the front-of-blog BGM) via OpenRouter using Google Lyria 3 Pro.
Animate a still image into a short video via OpenRouter. Default model google/veo-3.1-fast.
Generate images via OpenRouter. Default model openai/gpt-5.4-image-2; any image-modality model can be passed via --model.
Craft the COVER — the one animated <video> hero that opens the page above the headline, built from a Scout-verified still (or an abstract generated still) via a deterministic source ladder. The cover is animated by DEFAULT; a static <img> is only the recorded last rung. Owns the hero video + its static furniture; the Interaction Engineer owns the pick-a-side overlay, the Cinematographer stages the cover as cin_00. Reuses data-des=des_hero_video (no new provenance prefix). Outputs hero.json + assets/teaser.webm + _web.mp4 + .jpg poster.
Front stage for /data2story-pro when the reader has no dataset — only a vague idea. Converges the idea into a concrete, data-backed topic through a sparring-partner dialogue (anti-sycophantic, feasibility-pressure-tested), then acquires a REAL dataset through find-data, with a user checkpoint after each. Returns a validated DATA_DIR for the main pipeline. Not a newsroom role — runs upstream of Detective, before any dataset exists. Real data only; never a reason to synthesize data.
Research external context for a dataset — domain background, history, related studies, and why this data matters. Outputs detective.json (structured findings with det_xx IDs) before any analysis begins.
Generate videos via OpenRouter using ByteDance Seedance 2.0.
Claude Skills — questions
Answers built from the skills we actually parsed.