skills

Look up any arxiv paper on alphaxiv.org to get a structured AI-generated overview. Use this skill whenever the user shares an arxiv URL (arxiv.org/abs/...), an arxiv paper ID (e.g. 2401.12345), an alphaxiv URL, or asks you to explain, summarize, or analyze a research paper by ID or link. This is faster and more reliable than reading a raw PDF. Trigger proactively even when the user says things like "can you read this paper", "summarize this arxiv link", or "what does this paper say".

Use this skill whenever you need to elicit context, preferences, constraints, or choices from the user before proceeding. This is the default mechanism for all user-facing elicitation. The skill picks the best available tool, choosing a rich inline form (with cards, free-text, file uploads, multi-select, dates, sliders, and multi-round flows) when the visualizer is loaded, falling back to ask_user_input_v0 for choice questions, and prose for everything else. Trigger this any time you would otherwise ask a clarifying question in prose, call ask_user_input_v0 directly, or write a numbered list of questions for the user. This includes requests where intent is underspecified, decisions between options, gathering preferences for a recommendation, confirming assumptions before producing a deliverable, or collecting form-like data. Replaces ask_user_input_v0 as the default elicitation mechanism.

Build Python workflows and agents using the Claude Agent SDK (claude-agent-sdk). Use this skill whenever the user wants to create, design, or debug agent workflows, multi-agent systems, automation scripts, or any Python code that uses the Claude Agent SDK (formerly Claude Code SDK). Trigger on mentions of: "agent sdk", "claude-agent-sdk", "claude_agent_sdk", "query()", "ClaudeSDKClient", "subagents", "AgentDefinition", "agent workflow", "multi-agent", "claude code sdk", or any request to build an autonomous agent that reads files, runs commands, edits code, or orchestrates tasks programmatically. Also trigger when the user asks about architecting agent systems, choosing between SDK patterns, or integrating custom tools/hooks/MCP servers with the Agent SDK. Even if the user just says "build me an agent" or "automate this with Claude", use this skill.

Help the user understand and internalize complex topics through a structured pedagogical approach calibrated to how they learn. Triggers on "explain X to me", "help me understand Y", "what is Z and why does it matter", "break down this concept", "teach me about W", "how does X work", "walk me through Y", "I keep hearing about X but do not really get it", "what is the intuition behind X", or any request where the user wants to build a mental model of something they do not yet understand. Also trigger when the user asks conceptual "why" questions about systems, architectures, or paradigms. Do NOT trigger on simple factual lookups ("when was X founded"), summarization requests, problem-solving or decision-making (use the solve skill), or requests where the user already understands the topic and wants information organized.

Multi-pass self-review of work Claude just produced or modified, looking for bugs, errors, inconsistencies, and confusion. Trigger whenever the user says "fresh eyes review", "read carefully over what you just wrote", "look for any obvious bugs/errors/issues/problems", "review what you just modified", or "critically examine" what was just produced. Also trigger on prompts asking Claude to review work for typos, ambiguous wording, missing context, conflicting information, output-format inconsistencies, or logical/structural bugs. Also use proactively after producing a substantial deliverable (code file, document, slide deck, skill, spec, prompt). The skill runs a structured sequence of focused passes (surface, logic, cross-reference, regression) rather than one generic pass, because LLM self-critique misses different categories of issues on different passes. Do not use for trivial outputs like a one-sentence answer or a single-line edit.

Advanced Google search operator syntax for precise web searches. Read this skill BEFORE calling web_search when you need to restrict results to specific sites/domains, exclude terms or sites, find specific file types, filter by date range, find terms in proximity, or when basic keywords would return too many irrelevant results. Proactively use operators to improve search precision.

Design a codebase to work well with AI coding agents (Claude Code, Codex, Cursor, Aider). Use this skill when the user is starting or retrofitting a repository for agents, asking what AGENTS.md or CLAUDE.md should contain, designing project structure for AI assistance, asking about harness engineering, saying an agent keeps making the same mistake, asking how to make agents reliable on long-running tasks, or trying to make their codebase legible to coding agents. Trigger on indirect phrasings too. Examples include "set up my project for Claude Code", "make my repo agent-friendly", "stop my agent from doing X", "best architecture for AI-generated code", or "the agent keeps reinventing utilities". Use proactively when the user describes building a project with agent assistance and has not yet thought about scaffolding, AGENTS.md, custom linters, hooks, or feedback loops.

Apply post-2024 state-of-the-art principles for designing human-centered AI systems. Use this skill whenever the user is designing, specing, reviewing, critiquing, or planning an AI feature or product; asking principled questions about AI UX (errors, hallucinations, trust calibration, explainability, onboarding, anthropomorphism, agentic autonomy, citations, content provenance); working through tradeoffs around automation level or human-in-the-loop oversight; asking "should I use AI for this"; or evaluating designs against frameworks like Google PAIR, Microsoft HAX, Apple HIG, IBM, Stanford CS 247A, CMU 05-318, or recent CHI and UIST work. Trigger proactively even when the user does not explicitly ask for design principles, as long as they are working on an AI product question. Do NOT use for ML model selection, prompt engineering tuning detached from product design, or general AI ethics discussions detached from product context.

Remove signs of AI-generated writing from text. Use when editing or reviewing text to make it sound more natural and human-written. Based on Wikipedia's comprehensive "Signs of AI writing" guide. Detects and fixes patterns including: inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary words, negative parallelisms, and excessive conjunctive phrases.

Improve, clean up, or organize meeting notes, with or without an accompanying transcript. Triggers on requests like "improve my notes", "clean up my notes", "help me organize these meeting notes", "improve notes from transcript", or when a user provides raw meeting notes and/or a transcript and asks for cleanup. Also triggers when user provides notes and asks to merge them with a transcript or recording output.

Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get interviewed on their design, or mentions 'interview me'. Works across any domain: technical architecture, product strategy, writing, personal decisions, or anything complex that requires more context. Always triggers on 'interview me', 'stress-test this', 'poke holes in this', or 'challenge my thinking'.

Deep research methodology for investigating people, companies/organizations, and topics/events using publicly accessible sources. Use when research requires going beyond simple web search - finding connected information across multiple sources, verifying claims, building comprehensive profiles, or investigating entities systematically. Triggers on "investigate", "research [person/company]", "find information about", "background check", "verify", "deep dive", "OSINT", "find out everything about", or when a single search would be insufficient to answer the query.

Vision-based PDF reading that preserves layout, figures, tables, and visual formatting. Use this skill when reading or analyzing PDF content where visual elements matter - documents with charts, diagrams, complex tables, multi-column layouts, forms, or any PDF where text extraction would lose important information. Converts PDF pages to images and uses vision to read them, providing richer understanding than text extraction alone.

Audit whether an implementation matches its original plan by performing a blind reconstruction of the plan from code, then diffing it against the original. Use this skill whenever the user wants to verify code matches a plan, check for scope creep or missing functionality, compare what was built versus what was specified, audit Claude Code output against PLAN.md or SPEC.md, or invokes /plan-audit. Trigger on phrases like "audit my plan", "check if my code matches the plan", "did this implement everything", "any scope creep", "verify implementation", "review against spec", "is my code complete", or any request comparing built code to a planning document. Use this skill even when the user does not explicitly say "audit" but describes wanting to compare implementation against intended design.

Run the Qwen 3.5 0.8B language model locally in the Claude sandbox for inference and experimentation. Use this skill whenever the user asks to run Qwen locally, test a prompt on Qwen, run a small LLM, benchmark a small model, compare Claude's response to Qwen's, test Qwen 3.5 0.8B, run inference on a local model, experiment with a tiny language model, or any request involving running, testing, or benchmarking Qwen or small language models in the sandbox. Also trigger when the user wants to study calibration, hallucination, or reasoning differences between a small model and Claude, or asks "what would a small model say about this." Even casual mentions of "run Qwen" or "test on Qwen" should trigger this skill.

Structured problem-solving using consulting frameworks, for any domain. Triggers on vague challenges, decisions, diagnoses, prioritization, or strategy. Use whenever the user describes a problem and seems stuck, asks "should I...", "how do I prioritize...", "something is broken", "help me think through...", or any situation where they need structure, not just information. Also trigger on option comparisons, failure diagnoses, strategy planning, or figuring out where to start. Works across business, engineering, personal, team, career, health, creative, academic, and financial domains.

Iterative image manipulation for enhanced visual understanding. Use when you cannot extract information from an image due to orientation, scale, quality, or format issues. Provides rotation, cropping, scaling, contrast enhancement, and more. Trigger this skill whenever image content is unreadable, too small, rotated, low-contrast, or otherwise obstructed, even if the user does not explicitly ask for image processing.

Extract and analyze content from YouTube videos via transcription, subtitles, and visual frame analysis. Use this skill whenever the user shares a YouTube URL or asks to watch, listen to, summarize, analyze, transcribe, or extract information from a YouTube video. Also trigger on phrases like "what does this video say", "summarize this video", "watch this for me", "what happens in this video", or any request involving a youtube.com or youtu.be link. Trigger proactively even if the user just pastes a YouTube URL without explicit instructions -- assume they want the content analyzed.