Skills
97,501 packaged capabilities and playbooks for agents. Install any of them with one command.
Agent workflows
Reduce Claude Code token burn with proactive compacting, task-scoped sessions, checkpoint files, tiered docs, and relevance-based context pruning. Use when a user explicitly asks to reduce context usage/cost, run token benchmarks, or optimize CLAUDE.md. Do not use to expand product scope or skip correctness checks
Practical guide to reduce token consumption and lower AI costs for both Claude Code users and developers calling the Anthropic API directly. Covers file organization, context management, strategic model selection, prompt caching, Batch API, effort tuning, and prompt architecture. Backed by real experiment data. Use when user mentions "optimize tokens", "reduce costs", "Claude is slow", "too many tokens", "token budget", "context window full", "organize codebase for AI", "reduce token consumption", "prompt caching", "cache_control", "batch API", "cache broken", "API costs", or "thinking budget". Do NOT use for general coding questions, debugging, or performance optimization unrelated to AI token usage.
Save and mutate workflow state — init a workflow, update fields, transition phases, and capture a structured handoff before context exhaustion. Use when the user says 'save progress', 'checkpoint', 'I need to stop', or runs /checkpoint. Writes current workflow phase, task progress, and artifacts to the event store for later resumption. Do NOT use for restoring/reading state after a session break (that is /rehydrate) or for workflow initialization framing handled by ideate/debug/refactor commands.
An intelligent multi-store memory system with human-like encoding, consolidation, decay, and recall. Use when setting up agent memory, configuring remember/forget triggers, enabling sleep-time reflection, building knowledge graphs, or adding audit trails. Replaces basic flat-file memory with a cognitive architecture featuring episodic, semantic, procedural, and core memory stores. Supports multi-agent systems with shared-read, gated-write access. Includes philosophical meta-reflection that deepens understanding over time. Covers MEMORY.md, episode logging, entity graphs, decay scoring, reflection cycles, evolution tracking, and system-wide audits.
Guide for creating Claude Code skills following Anthropic's official best practices. Use when user wants to create a new skill, build a skill, write SKILL.md, update an existing skill, or needs skill creation guidelines. Provides structure, frontmatter fields, naming conventions, and new features like dynamic context injection and subagent execution.
Dispatch implementation tasks to agent teammates in git worktrees from a prepared capsule, then settle the batch. Triggers: 'delegate', 'dispatch tasks', 'assign work', or /delegate. Compiles the batch with prepare, spawns teammates, submits their results with settle. Supports --fixes flag. Do NOT use for single-file changes or polish-track refactors.
Collaborative design exploration that authors the Design & Rationale section of the one unified docs/specs/ spec. Triggers: 'brainstorm', 'ideate', 'explore options', or /ideate. Default is a one-pass design preamble; the deep rung gates the 2-3 approach divergent loop and the discover bridge. Do NOT use for task decomposition (that is /plan, which adds Decomposition to the same doc) or code review.
Plan work as vertical slices or an explicitly selected mechanism-reduction program in small, known-good increments, with independent pull requests or an optional stacked-PR topology across one or more ordered slices. Use when starting significant work, turning already-split stories into implementation plans, planning PRs, or sequencing complex tasks. For a mechanism-reduction program, use reduce-system-complexity first to define the conserved contract, terminal mechanism-removal state, and behavior/mechanism gates; planning then sequences it. If the input is a broad story, epic, feature idea, or backlog item that still needs product slicing, use story-splitting first; if a slice needs review layers or later slices should start on an evolving baseline before lower PRs merge, also use stack-pull-requests
Decompose dense codebase-wide, multi-document, PDF, and aggregation work even when the input fits the context window, following Recursive Language Models (Zhang, Kraska, Khattab, 2025). Use when the user asks to analyse all files, a whole repo, all docs, large PDFs, or to aggregate or multi-hop across scattered sources. Skip one file, one function, a single needle, or a one-page PDF conversion. Triggers: long context, context rot, large codebase, many files, all files, big document, multi-document, PDF, aggregate, summarize everything, codebase-wide, multi-hop, recursive, sub-agents, map-reduce.
Run `@metaharness/darwin evolve <repo>` to mutate a harness's seven policy surfaces (planner/contextBuilder/reviewer/retryPolicy/toolPolicy/memoryPolicy/scorePolicy), sandbox-score each variant, and promote only measured wins. The model is frozen; the harness evolves. Closes the loop ADR-150 opens (score+genome describe; evolve changes). Degrades gracefully when @metaharness/darwin is absent (ADR-150 + ADR-153 architectural constraints).
Manage agent memory through daily logs, session preservation, and knowledge extraction. Use when (1) logging work at end of day, (2) preserving context before /new or /reset, (3) extracting patterns from daily logs to MEMORY.md, (4) searching past decisions and learnings, (5) organizing knowledge for long-term retention. Essential for continuous improvement and avoiding repeated mistakes.
Lightweight workflow for straightforward changes — plan → implement → optional PR. Direct-commit by default; synthesize is opt-in via synthesisPolicy or a runtime request_synthesize event. Use for trivial fixes, config tweaks, single-file changes, or exploratory work that doesn't warrant subagent dispatch or two-stage review. Triggers: 'oneshot', 'quick fix', 'small change', or oneshot.
Push the git worktrees your workflow actually creates into Paseo's workspace registry, because Paseo discovers nothing on its own. Use when the user says "paseokit", "sync my worktrees to paseo", "sync all my projects to paseo", "my paseo sidebar is missing worktrees", "paseo doesn't show my worktree", or "where should paseo put new worktrees". It also reaps the finished work and the dead rows: "clean up the worktrees whose PRs merged", "clean up my paseo workspaces", "my paseo sidebar is full of dead entries".
Decompose the unified docs/specs/ artifact into parallelizable tasks — the ## Decomposition section of the same document whose ## Design & Rationale section holds the DR-N source. Triggers: 'plan implementation', 'create tasks from spec', or /plan. Applies the verification ladder: verification depth matches each task's blast radius — static analysis for low-risk tasks, scoped tests plus a kill-probe for medium, the integration suite on top for high-risk surfaces (judged test-after, not test-first ordering). Auto-chained from /ideate, or run directly to author the whole unified spec at thin/standard depth. Do NOT use for brainstorming, debugging, or code review.
Coordinate multiple DeepSeek Harness plugin Skills across inspection, migration, runtime debugging, heavy dependencies, testing, naming, and release. Use for a guided lifecycle workflow, a pre-run capability menu, or several DSH plugin operations with one status report. Preserve the user's selected scope and existing authorization across stages
Interactively prune stale non-terminal workflows from the pipeline. Use when the user says 'prune workflows', 'clean stale workflows', 'pipeline cleanup', or runs /prune. Runs a dry-run preview, displays candidates with staleness and safeguard skips, prompts the user to proceed/abort/force, then bulk-cancels approved workflows with a workflow.pruned audit event. Safeguards skip workflows with open PRs or recent commits unless force is set.
Restore and read workflow state after a context break — re-inject workflow phase, task progress, and behavioral guidance into the current session, reconcile state against git reality, and verify whether a workflow exists. Use when the user says 'resume', 'rehydrate', 'where were we', or runs /rehydrate, or when the agent has drifted after context compaction. Do NOT use for saving or mutating state (that is /checkpoint).
Orchestrates a full SEO + AI-search audit — acquires the site with the bundled scripts (crawl/snapshot), detects platform and vertical, runs the deterministic checks, dispatches the read-only specialist subagents in parallel with a dispatch envelope, merges their findings into a persisted report, and presents the two scores. Invoked by the `audit` and `geo` commands; not called directly.
Create pull request from completed feature branch using GitHub-native stacked PRs. Use when the user says 'create PR', 'submit for review', 'synthesize', or runs /synthesize. Validates branch readiness, creates PR with structured description, and manages merge queue. Do NOT use before review phase completes. Not for draft PRs.
Best practices for agents managing PostHog skills via the MCP `skill-*` tools — how to discover, read, create, update, and refactor skills efficiently, especially large skills with many bundled files. Use whenever you are about to call any `skill-*` tool, asked to author or edit a shared skill, or troubleshoot why a skill write was rejected. Also use when asked what skills are good for or how to shape one (project hub, runbook, catalog, handover, daily driver). Pairs with `skills-store` (which covers the raw tool surface) by adding the decision-tree, efficiency, pattern, and pitfall guidance
Use when you need to generate a checklist document with an embedded commands inventory, following the embedded template exactly and producing INVENTORY-COMMANDS-JAVA.md in the project root. This should trigger for requests such as Create embedded commands inventory checklist; Generate INVENTORY-COMMANDS-JAVA.md; Use @001-commands-inventory; Inventory embedded Java project commands; List command files bundled for Java agents. Part of Plinth Toolkit
Use when measuring whether an agent works: eval suites, trajectory judging, traces as regression fixtures, a calibrated LLM judge, and release gates. Triggers: “agent eval” / “эвалы агента”, “LLM judge” / “LLM-судья”. Not for ordinary unit tests or model benchmarks; the loop is agent-orchestrator.
Queries local analytics across OrchestKit projects for agent usage, skill frequency, hook timing, team activity, session replay, cost estimation, and model delegation trends. Privacy-safe with hashed project IDs. Supports time-range filtering and comparative analysis. Use when reviewing performance, estimating costs, or understanding usage patterns.
This skill provides guidance for creating agents and applications with the GitHub Copilot SDK. It should be used when the user wants to create, modify, or work on software that uses the GitHub Copilot SDK in TypeScript, Python, Go, or .NET. The skill covers SDK usage patterns, CLI configuration, custom tools, MCP servers, and custom agents.
Doctor Strange — causal sand-table simulation via parallel universe subagents. Models actors, incentives, constraints, feedback loops, state changes, and failure modes before rendering how a future event may unfold step by step, like a human mentally rehearsing multiple futures. Stores simulations as persistent memory for later recall. TRIGGER when: user explicitly asks to simulate / rehearse / play out a scenario; user says "推演", "模拟", "预演", "imagine", "what if", "run through", "play this out", "what could go wrong"; user faces a high-stakes upcoming decision and is uncertain how it will unfold. DO NOT TRIGGER when: user wants factual lookup or research; user wants analysis of a past event (use regular memory); user wants a simple recommendation without simulation; user is debugging code or doing technical work unrelated to decision-making. Three modes: SIMULATE (run a new forward simulation), RECALL (surface past simulations as soft priors), MANAGE (list/void/re-run stored simulations).
Run `@metaharness/darwin security bench` (upstream "Darwin Shield" / ADR-155)—evolve a champion security-detection harness against a 10-vuln / 9-decoy corpus and grade it on TPR/FPR/patch-pass/repro/unsafe against four baselines (B0 static, B1 LLM-single-pass, B2 fixed-agent, B3 Darwin-champion). Closest reference implementation for ruflo's own ADR-155 nightly self-learning security harness (PR #2417). Degrades gracefully when @metaharness/darwin is absent.
Guide for exposing PostHog product endpoints as MCP tools. Use when creating new or updating API endpoints, adding MCP tool definitions, scaffolding YAML configs, or writing serializers with good descriptions. Covers the full pipeline from Django serializer to generated TypeScript tool handler.
AWS-curated copy-paste prompts for AI coding agents (MVP scaffolding, RAG chatbot with Claude on Bedrock, security baseline evaluation, cost anomaly detection, GPU quota requests, EKS deployment, Well-Architected review, etc.) plus downloadable installable agents (Multi-Account Transition Advisor, Bill Shock Preventer, Service Quota Agent, Bedrock Model Availability Agent, AWS DB Advisor). Use when the user asks for a prompt to do X on AWS, wants an installable agent for multi-account / cost monitoring / quota management / Bedrock model availability / database selection, or asks how to use AWS prompts. For migration intent (GCP to AWS, Azure to AWS, OpenAI/Gemini to Bedrock), route to the migration skills in this plugin (`gcp-to-aws`, `azure-to-aws`, `heroku-to-aws`, `llm-to-bedrock`). Do not use for: factual AWS Activate / programs / credits questions, learn articles, sample architectures, or for prompts that are not in the bundled `references/prompt-library/` tree.
Captures learnings, errors, and corrections to enable continuous improvement. Use when: (1) A command or operation fails unexpectedly, (2) User corrects Claude ('No, that's wrong...', 'Actually...'), (3) User requests a capability that doesn't exist, (4) An external API or tool fails, (5) Claude realizes its knowledge is outdated or incorrect, (6) A better approach is discovered for a recurring task. Also review learnings before major tasks.
Use when you need to install the embedded project commands into command directories (.github/commands, .claude/commands, .cursor/command, .codex/commands), selecting the destination interactively and copying the embedded command definitions from project assets. This should trigger for requests such as Install embedded commands; Bootstrap .cursor/command; Bootstrap .claude/commands; Copy project commands; Install project command suite. Part of Plinth Toolkit
Use when you need to install the embedded robot agents into .github/agents, .claude/agents, .cursor/agents, or .codex/agents, interactively selecting the destination and copying the embedded agent definitions from project assets. Trigger for requests such as Install embedded agents; Bootstrap .github/agents; Bootstrap .cursor/agents; Bootstrap .claude/agents; Bootstrap .codex/agents; Copy robot agents. Part of Plinth Toolkit
SUPERSEDED by /claudex-loop (formerly /crucible; its docs-aware mode covers this variant) — prefer that skill unless you explicitly want this one. Two-act plan hardening with living documentation. ACT 1 (you ↔ Claude) — Claude interviews you relentlessly about a plan, one question at a time, challenging it against your project's existing domain model and glossary (CONTEXT.md), sharpening fuzzy terms, stress-testing with concrete scenarios, cross-referencing code, and updating CONTEXT.md + ADRs inline as decisions crystallise. ACT 2 (Claude ↔ Codex) — Claude writes the locked plan to PLAN.md and OpenAI Codex adversarially reviews it in a read-only sandbox (VERDICT:APPROVED/REVISE), Claude revises and re-submits to the SAME Codex session until APPROVED or a MAX_ROUNDS cap, then you sign off before any code. Use when the user says "/grill-with-docs-codex", "grill me against the docs then have codex review", "stress-test this against our domain model then get a second model on it", or is about to build something high-stakes in a project with established terminology/ADRs and wants alignment, documentation, AND a cross-model sanity check. Builds on Matt Pocock's grill-with-docs (MIT). NOT for reviewing already-written code (use /codex:review) and NOT for trivial changes
Manage `@metaharness/darwin` bench suites — `bench create <repo>` scaffolds a JSON suite from a repo's test corpus; `bench verify <suite.json>` checks suite well-formedness. Bench suites are the fixed evaluation corpora that `harness-evolve --bench <suite.json>` scores variants against, decoupling evolution from the repo's natural tests. Degrades gracefully when @metaharness/darwin is absent.
Guide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).
Comprehensive security review framework for AI agents. Covers skill/MCP installation, GitHub repos, URLs/documents, on-chain addresses, products/services, and social shares. Built from real-world attack patterns and incident response experience.
Maintains a repo-local task forest or task DAG for the current workspace. Use when the user asks to initialize or update a task forest, close a session, summarize evolving project work, align a request with a global goal, track progress/history/deviations/todos, save or apply a task proposal, or export the client-readable task-forest HTML. Do not use for executing the tracked tasks themselves or for generic HTML work.
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.