Skills
97,222 packaged capabilities and playbooks for agents. Install any of them with one command.
Agent workflows
Sharpen the prompt before you send it, whether the one-shot instruction you're about to hand a coding agent or the system prompt your application ships. Use when the user says "optimize this prompt", "what's wrong with this prompt", "write or rewrite the system prompt my app ships", or "/promptkit".
Scaffold a custom AI agent harness via `metaharness new <name> --template <id> --host <id>`. Defaults to DRY-RUN (no writes) unless --confirm is passed. Refuses to write to the calling repo root or anywhere inside it. Honors ADR-150 architectural constraint + ruflo's "destructive-action confirmation" pattern.
Memory retrieval internals: knowledge graph orchestration with entity extraction, natural language query parsing, deduplication (>85% similarity), and cross-reference boosting over unified recency, relevance, and authority ranking. Use when designing or debugging how memory search itself works. Everyday lookups belong to memory; entry storage to remember; consolidation to dream.
Run existing ShinkaEvolve tasks with the `shinka_run` CLI from a task directory (`evaluate.py` + `initial.<ext>`). Use when an agent needs to launch async evolution runs quickly with required `--results_dir`, generation count, and strict namespaced keyword overrides.
Use when the user wants their Claude agent to improve itself based on past usage, asks about a nightly/offline “sleep” or “dream” cycle, memory/skill consolidation, or says things like “make my agent better the more I use it,” “review my past sessions,” “learn my preferences,” “consolidate what you learned,” “run the sleep cycle,” or wants to schedule background self-optimization. Drives the skillopt_sleep engine: harvest past sessions -> mine recurring tasks -> replay through a selected backend -> consolidate validated CLAUDE.md/SKILL.md behind a held-out gate.
Wire a semantic layer into a nao agent so that metric queries are routed through a single source of truth. Supports dbt MetricFlow (dbt Cloud with Semantic Layer), Snowflake (views or semantic views via MCP), an in-house nao YAML semantic layer, or other tools (via MCP discovery). Installs the right MCP server, updates RULES.md to route metric queries through the semantic layer, and (for the nao YAML option) generates starter metric files. Use after a first round of tests has shown the agent struggling with metric reliability. Do not use for raw rule writing (write-context-rules) or first-time setup (setup-context).
Build and edit Dify workflow DSL files. Use when creating new Dify workflows from scratch, modifying existing workflows (adding/removing nodes, changing connections), validating workflow structure, or designing LLM-based automation flows with code execution, conditional logic, loops, and error handling.
Drive a pull request's change end to end: decide it's flag-worthy, create the guarding flag, wire the new code path behind it on the PR branch, and record an automated release so the change ships safely when the PR merges. A portable orchestrator that composes should-flag-change, launchdarkly-flag-create, and flag-release. Keywords: flag a PR, wrap change in a flag, dark launch, kill switch, auto-release, automated rollout, end-to-end flag workflow.
One-command drift detection. Composes audit-list + oia-audit + audit-trend into a single primitive — finds the most recent audit in `metaharness-audit` namespace, runs a fresh audit against the current repo, diffs them via ADR-152 §3.1 similarity, and alerts when structural distance crosses `--threshold`. Iter 53 of ADR-150 deep integration.
ADR-152 — weighted similarity between two harness fingerprints (genome + score JSON). Returns overall score in [0,1] plus per-component breakdown (cosine over 9 numerics, categorical agreement over 4 enums, jaccard over agent_topology). Unblocks ADR-151 §3.2 Recommender, §3.3 Drift Detection, §3.5 Plugin Compat. Pure TS, no `@metaharness/*` dep — preserves ADR-150's four architectural constraints.
Analyzes, documents, and maintains a living `.agent-docs/` knowledge base for Unity projects targeting Meta VR and Horizon OS. Use when the user asks to scan project structure, explain how a Unity system works, or update project docs after structural changes.
Run AI apps via inference.sh CLI - image generation, video creation, LLMs, search, 3D, Twitter automation. Models: FLUX, Veo, Gemini, Grok, Claude, Seedance, OmniHuman, Tavily, Exa, OpenRouter, and many more. Use when running AI apps, generating images/videos, calling LLMs, web search, or automating Twitter. Triggers: inference.sh, infsh, ai model, run ai, serverless ai, ai api, flux, veo, claude api, image generation, video generation, openrouter, tavily, exa search, twitter api, grok
Audit any second brain, notes folder, or agent memory for facts that have quietly stopped being true, then fix the worst one so it stops recurring. Works on a wiki, a single notes file, daily notes, or a non-markdown tool, and adapts the fix to whichever it finds. Checks every current-sounding claim in whatever the agent loads each session against the freshest evidence, and separates contradicted claims from unsupported ones. Use when an assistant gives an outdated answer, when notes or memory files may be stale, when a vault needs checking for contradictions, or when someone asks how to stop a second brain from rotting, mentions memory rot, or asks about state versus event
Operate a reusable technical book manuscript workspace with writing structure, reader persona SSOT, review rules, and optional Markdown to Re:VIEW/PDF support. Use when organizing a book manuscript repo, defining or reviewing target readers, standardizing chapter/section files, setting writing/review agents, handling publisher proof comments, or assessing an existing writing workspace. Triggers on book writing workspace, reader persona, reader audience, 読者ペルソナ, persona SSOT, technical book project, 執筆ワークスペース, manuscript workflow, Re:VIEW workspace, proof PDF, annotated proof, publisher proof comments, 校正, 校正注釈, and 初校.
Build RAG pipelines with Exa.ai for real-time web retrieval. Use when building retrieval-augmented generation, integrating Exa with LangChain, LlamaIndex, Vercel AI SDK, or implementing AI agents with web search capabilities. Triggers on: RAG pipeline, retrieval augmented generation, Exa LangChain, Exa LlamaIndex, ExaSearchRetriever, ExaSearchResults, Exa MCP, Exa tool calling, Claude tool use, AI agent web search, grounded generation, citation generation, fact checking, hallucination detection, OpenAI compatibility, chat completions.
Generate a phased delivery orchestration plan for creative-direction-driven work: which skills run when, what locks at which gate, how handoffs occur, and how the cadence is implemented in the team's tools (Jira, Linear, Notion, Figma, GitHub) with AI agents via MCP and CLI. Provides a temporal map of phases, lock points, handoff specs, gate definitions, and QA verification gates. Distinct from `creative-brief` (static project brief) and `creative-direction` (aesthetic direction). Triggers on integration orchestrator, delivery cadence, project orchestration, when does the brief lock, how to phase creative direction, design-development handoff, QA gate, agent-driven QA, and when a creative-direction project is starting, mid-flight with broken cadence, or integrating AI agents into PM workflow. Does NOT fire for project scoping (use `creative-brief`), aesthetic direction (use `creative-direction`), or general project management without a creative-direction throughline.
Review the current conversation and capture valuable knowledge — best practices, coding conventions, architecture decisions, workflows, and user feedback — into persistent memory (AGENTS.md) or reusable skills. Use when the user says: (1) remember this, (2) save what we learned, (3) update memory, (4) capture learnings
TanStack AI (alpha) provides provider-agnostic, type-safe chat with streaming for OpenAI, Anthropic, Gemini, and Ollama. Use for chat APIs, React/Solid frontends with useChat/ChatClient, isomorphic tools, tool approval flows, agent loops, multimodal inputs, or troubleshooting streaming and tool definitions.
Use xurl to access AI agent conversations via agents:// URIs. Invoke when the user gives an agents:// URI, a provider shorthand like provider/..., or a bare thread/session ID; mentions conversations, threads, or sessions from any AI coding agent; wants to search, read, summarize, compare, or continue agent threads; asks what they worked on, what an agent said, or references past agent interactions; or wants to delegate work to or start a conversation in another agent.
Deploys swarms of sub-agents for massive parallel data processing tasks. Unlike agent-army (which is for code changes), this is for DATA tasks -- processing 1000 documents, analyzing datasets, bulk content generation. Configurable swarm size, task distribution, result aggregation, progress tracking, and error recovery.
Capture each agent failure, bucket and count it deterministically, and when a failure class recurs past a threshold, scaffold a candidate rule/skill stub into the propose->gate flow so the same mistake becomes a guardrail instead of repeating. Hamel's create->label->fix->repeat flywheel, self-evolving. Trigger when the user says "log this failure", "record the mistake", "why does this keep happening", "turn failures into rules", "failure taxonomy", "what keeps breaking", or invokes /failure-flywheel. Also run --report periodically to see the top recurring failure classes.
Scaffold a new skill in both publish trees at once — skills/<name>/SKILL.md plus its byte-identical llmwiki mirror — via fdk/tools/new-skill.py, then print the exact commands to register it on the curated surfaces (LOOP_MAP, marketplace.json, AGENT.md/CLAUDE.md). Use when the user says "create a skill", "add a new skill", "scaffold a skill", "make a new skill", "new skill", or invokes /new-skill.
Evaluates which SaaS tools can be replaced with AI agents. Takes a list of current SaaS subscriptions with costs, assesses replacement feasibility, estimates build vs buy economics, identifies Claude+MCP alternatives, and generates a comprehensive replacement plan with priority matrix, ROI analysis, implementation timeline, and risk assessment
Self-reflection + Self-criticism + Self-learning + Self-organizing memory. Agent evaluates its own work, catches mistakes, and improves permanently. Use when (1) a command, tool, API, or operation fails; (2) the user corrects you or rejects your work; (3) you realize your knowledge is outdated or incorrect; (4) you discover a better approach; (5) the user explicitly installs or references the skill for the current task
Ask the user yes/no, pick-one or A/B/C design questions on a local review board and read back their answers, using the sign-off-axi CLI. Use when you need a yes/no decision, a choice between a few options, a design pick between mocks, or approval from the user before going on, want the user to review or sign off on a plan or on proof of finished work (screenshots or a recording), need to read back the answers the user already gave, must explain a concept, design or trade-off to the user as an interactive report, want to track work as tickets the user can follow, or need the user to sign in to a service
Run a Karpathy-style LLM council (3-stage multi-agent evaluation) on top of the existing orca orchestration: N seats independently answer a question, the answers are BLIND peer-ranked (identities stripped so a model can't favour its own), and a chairman synthesizes the final answer. The orchestration makes the model calls; harness/scripts/council.py does the deterministic protocol math (anonymization, mean-rank aggregation, dissent, anchor guard, transcript). Use when the user says "run a council", "llm-council", "panel of models", "blind peer-rank these answers", "ensemble + chairman", or invokes /council. Stage-4 is a MANDATORY HTML report, rendered inside council.py itself (offline, no CDN, no external skill): every `rank` writes a NEW versioned self-contained report (llmwiki/html/council/council-report-NNN-seed<seed>.html) — opinion cards (persona name + lens), a blind-vote table, and a dashboard. Isolated in try/except so a render bug never kills the core transcript.
Front-door on-demand cho phát triển framework HOẶC distill/author một skill — SELF-CONTAINED, chạy được ở BẤT KỲ project nào (không phụ thuộc file repo-local). Gọi khi đang sửa chính framework, hoặc đang viết/chưng cất một skill trong phiên của dự án khác. KHÔNG dùng cho dev tính năng dự án thường (đó là phần lớn phiên — ADR-004).
Deterministic guardrailed agent-loop driver — wrap any step with propose → deterministic-verify → (critique → revise) and enforce hard termination guards (max_iter, wall-clock budget, no-progress via state-hash, escalate-to-human). The control loop + guards + progress detection are deterministic; the LLM critique/revise step is the one quarantined adapter. Trigger when an agent fix-loop must not spin forever and must not be able to argue its way around a failing test.
Fast, accurate code search for AI agents using ~98% fewer tokens than grep+read. Indexes any local or remote repository in under a second (~250ms on CPU, no GPU or API key needed). Supports natural-language and symbol queries, semantic similar-code discovery, and MCP server integration for Claude Code, Codex, Cursor, and OpenCode. Python library available for programmatic use. Triggers on: semble, code search, semantic code search, semble search, token-efficient search, find code, code search mcp, agent code search, semble find-related, semble savings
Re-inject Synapse A2A initial instructions after context has been cleared. Use this skill when the agent has lost its identity, A2A communication protocol, or configuration after /clear or context reset, when synapse send/reply commands stop working, or when the agent no longer recognizes its SYNAPSE_AGENT_ID. Triggered by /synapse-reinst command.
Recovers decisions, previous fixes, research, and what followed a prompt from past AI conversations, with source evidence. Use when asked to “search past chats”, “we fixed this before”, “what followed this prompt”, or “why did the plan change”. Reads local Claude Code, Codex, Grok, and Cursor history plus explicit ChatGPT or Claude exports.
Root-causes AI agent failures from MLflow traces without requiring a code change, regression suite, scorer, or evaluation run. Use when the user asks to debug, investigate, explain, or determine why an agent produced an output. Prefer an existing matching trace; reproduce only when no usable trace exists.
Discover and install agent skills from the open ecosystem. Use IMMEDIATELY when the user asks “how do I do X,” “find a skill for X,” “is there a skill for X,” or “can you do X” for specialized capabilities, wants to extend agent functionality, mentions a domain they wish they had help with, or asks about tools/templates/workflows that might exist as installable skills.
Convert an existing codebase in the current working directory into a ShinkaEvolve task directory by snapshotting the relevant code, adding evolve blocks, and generating `evaluate.py` plus Shinka runner/config files. Use when the user wants to optimize existing code with Shinka instead of creating a brand-new task from a natural-language description.
Which skill owns this task. Invoke BEFORE doing by hand any of: .pdf read or fill, test plan, architecture improvement, second-model review, docs writing, greenfield page, prose cleanup, 「有沒有 skill」, 「用哪個」 — or any task not named in the kernel routing index. Even a 1% chance a skill covers it means invoke this first; it costs one read.
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.