Skills
96,981 packaged capabilities and playbooks for agents. Install any of them with one command.
Agent workflows
Cross-check a hard bug with independent diagnostic workers only when the user explicitly requests them; otherwise run one inline diagnosis. Use when a bug is high-stakes, has resisted a single diagnosis pass, or when the user wants to triangulate or cross-check before committing to a fix.
Use when the user is hitting Claude usage limits, burning through tokens fast, running a long conversation, or asks how to use Claude Code more efficiently. Triggers on phrases like "hit my limit", "running out of tokens", "usage limit", "save tokens", "reduce usage", "am I wasting tokens", "this chat is getting long", "which model should I use". Diagnoses the current session, runs a real token-usage report from local JSONL logs, and applies Dubi's 11 rules for reducing Claude usage.
Create or update repo-native project instructions for AI agents, such as AGENTS.md or CLAUDE.md. Use when initializing an AI-native repository protocol, documenting project overview, tech stack, repository structure, build/run/test commands, durable working rules, progressive disclosure, workflow artifact routing, organization skill routing, or converting repeated coding-session habits into durable agent instructions.
All network operations must be handled through this skill, including: searching, web scraping, post-login operations, dynamic page interaction, and more. Trigger scenarios: the user asks to search for information, view web page content, access sites requiring login, operate web interfaces, scrape social media content (Xiaohongshu, Weibo, Twitter, etc.), read dynamically rendered pages, and any network task requiring a real browser environment.
Create workflow-* skills by composing existing skills into end-to-end chains. Turns a user idea into a workflow_spec.md SSOT (via workflow-brainstorm), discovers available skills locally + from skills.sh, and generates a new workflow-<slug>/ skill package. Use when you want to design a new workflow, chain multiple skills into a flow, or turn scattered atomic skills into a resumable plan-then-confirm workflow.
Execute written implementation plans: first read and critically review the plan, then implement in small batches (default 3 tasks), produce verification evidence per batch and pause for feedback; must stop immediately and ask for help when blocked/tests fail/plan unclear. Trigger words: execute plan, implement plan, batch execution, follow the plan.
Ship Faster end-to-end workflow for small web apps (default: Next.js 16.1.1): idea/prototype → foundation gate → design-system.md → lightweight guardrails + docs → feature iteration → optional Supabase + Stripe → optional GitHub + Vercel deploy → optional AI-era SEO (sitemap/robots/llms.txt). Resumable, artifact-first under runs/ship-faster/ (or OpenSpec changes/). Trigger: ship/launch/deploy/production-ready MVP.
Independence-validated parallel fleet that runs each worker (claude -p, codex exec, or pi -p) in its own git worktree. Use when tasks touch non-overlapping files and you need merge-safe isolation (each worker on its own branch). For DAG-ordered one-shot workers with budgets, use dag-fleet. For headless iteration with a reviewer loop, use iterative-fleet.
A design charter for designing, implementing and procuring the harness of AI agents (loop, tool surface, context management, authorization, state, stop conditions, evaluation, observability). Use it for "design the harness", "should we build it ourselves, use an SDK or an off-the-shelf option", "what to put in tool/MCP responses, how much to include and how to truncate", "agent permissions and approval tiers", "compaction and multi-agent", "when real model responses differ from expectations or are unstable", and "the mechanics of verify, eval and LLM judge". On the repository side see agent-native-project-design; on Mastra see mastra-ai-architecture-rules.
A skill for designing, organizing and pruning the repository/project side (CLAUDE.md / AGENTS.md, Agent Skills, hooks, subagents) that runs on off-the-shelf AI agents and harnesses such as Claude Code / Codex. Use it for consultations like "design/organize CLAUDE.md", "I want to hand this work to an agent", "I want to create a skill", or "the agent doesn't follow instructions". For developing the harness/runtime itself, see agent-harness-engineering.
Use this skill when the user wants a human-readable project progress screen, agent work visualization, customer-facing status cockpit, immediate bootstrap progress before TASK-PLAN.md exists, task-plan-derived HTML progress page, user decision gates, or auto-updating UI/UX project status for senior-ui-ux-orchestrator workflows. It renders and validates progress views from project-request.json, workflow-log.jsonl, TASK-PLAN v2, progress-state.json, feedback artifacts, reports, and wiki decisions without becoming the source of truth.
Use this skill for any request whose goal is session continuity across agents or time: finding a pending handoff, resuming previous work, saving next-session context, wrapping up, or discarding a mistaken handoff. Trigger by semantic intent rather than exact wording.
Distill books, documents, folders, or URLs into compact, practical Agent Skills. Use when source material should become either one reusable skill package or a split set of focused skills, each with a concise SKILL.md plus detailed references and examples.
Records research provenance at the end of a coding or research session by scanning the conversation and writing decisions, experiments, dead ends, pivots, claims, and heuristics into the ara/ directory (exploration_tree.yaml, claims.md, heuristics.md, observations.yaml, session records), each tagged user, ai-suggested, ai-executed, or user-revised. Use when a task is finished and you want a session epilogue. Use when you need an auditable trace of how a project evolved. Use when decisions and dead ends need logging with provenance. Use when initializing an ara/ directory. Do not use during task execution, since it must run only after the request is complete.
Orchestrates end-to-end autonomous AI research projects using a two-loop architecture. The inner loop runs rapid experiment iterations with clear optimization targets. The outer loop synthesizes results, identifies patterns, and steers research direction. Routes to domain-specific skills for execution, supports continuous agent operation via Claude Code /loop and OpenClaw heartbeat, and produces research presentations and papers. Use when starting a research project, running autonomous experiments, or managing a multi-hypothesis research effort.
Baidu Drive file management — upload, download, transfer, share, search, move, copy, rename, and create folders. Also supports Agent memory backup and restore (kimiclaw/maxclaw/qclaw/openclaw). TRIGGER: The user mentions "百度网盘/bdpan/网盘/云盘/baidu drive/Baidu Drive" in connection with file operations, or mentions memory-related operations such as "备份记忆", "恢复记忆", or "查看记忆备份". DO NOT TRIGGER: Non-file-storage operations, use of other cloud-drive services, local memory organization/cleanup, or PPT generation (handled separately by the baidu-wenku-aippt skill).
Makes a task or prompt more ambitious using Felix Rieseberg's types of ambition, Altitude (aim at the outcome, the goal beyond the goal), Depth (work like the best expert in the field, from the raw data, checked by an outside judge) and Range (bring in other fields and do what their experts do). Writes the ambitious plan and a rewritten prompt, then does the work when it was handed a task. Use when the user types /be-ambitious, says "be more ambitious", "aim higher", "go deeper on this", "what's the goal beyond the goal", "what would a top expert do here", "raise the bar", or asks to make a prompt more ambitious
Brief the user on the current task so they can decide next steps without reloading the codebase. Use only when they explicitly ask for a briefing or catch-up on current state, the likely implementation approach, the relevant code or architecture, or the facts needed for a decision.
Show real token usage and estimated savings for the current session. Reads directly from the Claude Code session log — no AI estimation. Triggers on /caveman-stats. Output is injected by the mode-tracker hook; the model itself does not compute the numbers.
Creates and reviews CLAUDE.md configuration files for Claude Code. Applies HumanLayer guidelines including instruction budgets (~50 user-level, ~100 project-level), WHAT/WHY/HOW framework, and progressive disclosure. Identifies anti-patterns like using Claude as a linter for style rules.
Build Model Context Protocol (MCP) servers on Cloudflare Workers - the only platform with official remote MCP support. TypeScript-based with OAuth, Durable Objects, and WebSocket hibernation. Use when: deploying remote MCP servers, implementing OAuth (GitHub/Google), using dual transports (SSE/HTTP), or troubleshooting URL path mismatches, McpAgent exports, OAuth redirects, CORS issues.
Optimize Codex Plus/Pro usage before and during coding tasks. Use when the user wants to save Codex quota/credits, reduce token or context waste, choose a cheaper model/reasoning level, avoid unnecessary repo scans/tests/subagents, or plan an implementation with the smallest reliable execution footprint. Do not use when the user explicitly prioritizes maximum quality regardless of usage.
Resolve a consequential decision by developing genuinely independent, credible positions and choosing between them with explicit criteria. Use when a question has two or more defensible answers, when evidence is ambiguous or contested, or when a single position risks sycophantic or premature consensus. Not for ordinary drafting or format choices.
Distill Opus-level reasoning into optimized instructions for Haiku 5.5 (and Sonnet). Generates explicit, procedural prompts with n-shot examples that maximize smaller model performance on a given task. Use when the user says "down-skill", "distill for Haiku", "optimize for Haiku", "make this work on Haiku", "generate Haiku instructions", or needs to delegate a task to a smaller model with high reliability
Only use when the user explicitly invokes /execute (or $execute in Codex). Never auto-invoke. An orchestration skill that organizes and splits complex tasks and unifies the available delegation mechanisms to carry out a project. It focuses on overall organization and coordination rather than implementation details, sorts out task dependencies and delegates independent tasks in parallel. Use when explicitly triggered with "execute the task", "carry it out end to end" or "orchestrate it".
New to MCP?
From what the protocol is to installing your first server, these answers cover most first questions.
An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.