agent-skills

Use Claude's Code Execution tool to run Python in a sandboxed environment as part of a response — for calculation, data analysis, chart generation, and verification. Covers enabling, file upload, persistence across turns, and limitations. Use this skill when building features that need Claude to actually run code (not just write it), such as data analysis, math verification, or chart creation. Activate when: Claude code execution, Python sandbox, run code tool, data analysis agent, code interpreter, code_execution_20250522.

Build browser/desktop automation agents using Claude's Computer Use capability — screen-taking, clicking, typing. Covers the reference container, virtualization safety, task decomposition, and when to use computer-use vs API integration. Use this skill when building agents that operate GUIs (browsers, legacy apps), automating workflows without APIs, or QA/testing agents. Activate when: Claude computer use, browser automation, desktop agent, screen control, computer_20250124, click and type agent.

Use Claude's extended thinking (reasoning) mode effectively — budget tokens, interleaved thinking with tool use, when it helps, when it wastes tokens, and how to inspect the thinking trace. Use this skill when building reasoning-heavy features (math, code generation, multi-step planning), debugging why a model is shallow on hard problems, or deciding whether to enable thinking. Activate when: extended thinking, thinking tokens, budget_tokens, reasoning mode, interleaved thinking, thinking blocks.

Use Claude's 1M-token context window effectively — when to use it, how to structure inputs for recall, how to price it, and how to combine with prompt caching to keep it affordable. Use this skill when building apps that feed large codebases, long documents, or entire conversation histories to Claude, or when weighing 1M context vs RAG. Activate when: 1M context, long context, big context window, context vs RAG, Claude 1 million tokens, context-beta header.

Use Claude's Memory tool to give agents persistent cross-session memory stored in a client-side file directory. Covers setup, directory layout, reading/writing patterns, and when memory beats context-stuffing. Use this skill when building agents that need to remember across sessions (user preferences, project state, past decisions), or when moving from "stuff everything into context" to persistent memory. Activate when: Claude memory tool, agent memory, persistent memory, memory directory, cross-session state, long-term memory agent.

Use Claude's prompt caching with 5-minute and 1-hour TTLs to slash costs on repeated context — codebases, system prompts, long documents. Covers cache breakpoints, hit-rate optimization, and the common mistakes that silently disable caching. Use this skill when building apps with repeated large context, optimizing LLM spend, or debugging "why are my cache reads zero?" Activate when: prompt caching, cache_control, cache hit rate, 5 minute cache, 1 hour TTL cache, ephemeral cache, reduce Claude cost.

Build your first Claude Agent SDK agent — TypeScript or Python. Covers installation, minimal agent loop, tool definitions, and the difference between "roll-your-own" and Managed Agents. Use this skill when starting a new AI agent project, migrating from raw messages.create loops to the Agent SDK, or evaluating SDK vs direct API. Activate when: Claude Agent SDK, @anthropic-ai/claude-agent-sdk, build agent, agent SDK quickstart, agent tutorial.

Stream agent output correctly — text deltas, tool-use events, thinking deltas, progress indicators — without dropping events or blocking on long tool calls. Covers backpressure, error handling, and UX patterns. Use this skill when building agent UIs (chat, CLI), ensuring agents feel responsive, or debugging dropped/truncated streams. Activate when: agent streaming, stream events, text deltas, streaming UI, SSE agent, stream tool use.

Use Anthropic's Managed Agents API (/v1/agents, /v1/sessions) — server-side agent runtime that handles the tool loop, compaction, memory, and scaling for you. Covers when to pick Managed over SDK, request shape, and cost model. Use this skill when building production agents at scale, deciding between SDK vs Managed, or migrating from self-hosted agent loops. Activate when: Managed Agents API, /v1/agents, /v1/sessions, server-side agent, Anthropic agent runtime.

Manage Claude Agent SDK / Managed Agents sessions — creation, resumption, compaction, forking, and termination. Covers when to start fresh, when to resume, and how to handle context window pressure in long sessions. Use this skill when building multi-turn agents, debugging "my agent forgot the earlier context", or designing session retention policies. Activate when: agent session, session resume, conversation compaction, session fork, context window overflow, session lifecycle.

Delegate work to sub-agents via the Task/Agent tool — parallel research, isolated context windows, specialized expertise. Covers when sub-agents help vs hurt, prompt shape, and result handling. Use this skill when building agents that need to research in parallel, process independent work items, or isolate context-heavy sub-tasks. Activate when: sub-agents, Task tool, Agent tool, parallel agents, agent delegation, spawn agent, multi-agent.

Measure and optimize the cost/quality curve — which model, prompt, and settings give the best quality per dollar. Covers Pareto analysis, break-even thresholds, and when to spend more vs less. Use this skill when optimizing LLM spend, picking a default model for a feature, or deciding whether a premium model is worth it. Activate when: cost vs quality, model selection, eval cost, Pareto frontier, cheaper model, premium model tradeoff.

Design eval datasets that actually measure model quality — coverage, difficulty distribution, labeling consistency, and avoiding contamination. Covers sourcing, stratification, label quality, and when to generate vs curate. Use this skill when building a new eval set, realizing your current evals don't catch regressions, or labeling is inconsistent. Activate when: eval dataset, benchmark, test set, eval coverage, label quality, synthetic eval, dataset design.

Curate and maintain "golden set" eval items — the small, high-signal cases that must never regress. Covers selection criteria, review cadence, retiring stale items, and keeping the set sharp. Use this skill when building a sanity-check eval that runs on every PR, defending against silent quality drops, or your full eval takes too long to run in CI. Activate when: golden set, smoke test eval, canary eval, must-not-regress, eval sentinels, core eval.

Use an LLM as an evaluator for open-ended outputs — rubrics, pairwise comparison, calibration with human labels, bias mitigation. Covers when LLM-judge works, when it fails, and how to trust its scores. Use this skill when evaluating generative outputs at scale, building eval pipelines, or replacing expensive human review for non-critical judgments. Activate when: LLM as judge, LLM evaluator, automated evaluation, pairwise comparison, rubric evaluation, eval model.

Set up continuous regression evals so model/prompt/tool changes don't silently break existing behavior. Covers gating thresholds, CI integration, statistical significance, and response to regressions. Use this skill when deploying prompts to production, gating model upgrades, or noticing "it worked yesterday" in AI features. Activate when: regression eval, eval CI, prompt regression, model upgrade gate, eval threshold, eval alert.

Implement OAuth 2.1 + PKCE authentication for remote MCP servers, including dynamic client registration, token refresh, and scope design. Covers the 2025 MCP auth spec that Claude Desktop, Claude Code, and ChatGPT use. Use this skill when building a remote MCP server that needs per-user auth, debugging OAuth flows for MCP, or migrating a bearer-token MCP server to OAuth. Activate when: MCP OAuth, remote MCP auth, MCP authorization, PKCE, dynamic client registration, MCP 401.

Connect MCP servers to AI clients — Claude Desktop, Claude Code, Cursor, the Claude Agent SDK, and custom apps — including config file locations, auth, and debugging connection failures. Use this skill when the user wants to wire an MCP server into a client, troubleshoot "server not connecting" errors, or build a custom MCP client with the SDK. Activate when: MCP client, mcp.json, claude_desktop_config.json, connect MCP, MCP not working, list MCP tools, programmatic MCP client.

Use MCP Resources correctly — the read-only, URI-addressable data primitive — and know when to pick resources, tools, or prompts. Covers templated URIs, subscriptions, and common mistakes. Use this skill when designing MCP servers that expose data, deciding between tool vs resource, or implementing resource subscriptions for live data. Activate when: MCP resource, resource template, resource vs tool, subscribe resource, MCP URI schema.

Secure MCP servers against prompt injection, tool abuse, excessive permission, and data exfiltration. Covers per-tool scopes, rate limiting, audit logging, and sandbox patterns for shell-adjacent tools. Use this skill when deploying an MCP server to production, handling untrusted agents, or reviewing an MCP server for security issues. Activate when: MCP security, MCP prompt injection, tool sandbox, MCP audit log, MCP rate limit, tool abuse, MCP threat model.

Build production-quality MCP (Model Context Protocol) servers that expose tools, resources, and prompts to AI clients like Claude Desktop, Claude Code, Cursor, and the Claude Agent SDK. Use this skill when the user wants to build an MCP server, expose internal tooling to an AI agent, wrap an API for agents, or publish a reusable MCP server to a registry. Activate when: MCP server, Model Context Protocol, expose tools to Claude, @modelcontextprotocol/sdk, mcp.json, stdio server, HTTP MCP server.

Design MCP tool schemas, names, and descriptions that AI agents actually pick correctly and use without hand-holding. Covers the anti-patterns that make agents loop, pick wrong tools, or hallucinate arguments. Use this skill when designing or reviewing MCP tools, debugging "the agent isn't using my tool", or pruning a bloated tool surface. Activate when: MCP tool design, tool description, agent picks wrong tool, too many tools, tool schema, tool naming.

Choose between MCP transports — stdio for local processes, Streamable HTTP for remote servers, SSE for legacy — and implement each correctly with reconnection, backpressure, and session handling. Use this skill when deciding transport for a new MCP server, migrating SSE to Streamable HTTP, or debugging transport-level issues (connection drops, buffering, session loss). Activate when: MCP transport, stdio vs HTTP, Streamable HTTP, MCP SSE, MCP reconnection, MCP session.

Use this skill when managing AI agent costs. Activate when the user needs to control token usage, implement cost limits for agents, optimize LLM spending, track agent costs, or prevent runaway API bills in agent systems.

Use this skill when implementing fallback strategies for LLM applications. Activate when the user needs graceful degradation for AI services, multi-provider failover, handling LLM outages, or building resilient AI systems.

Use this skill when implementing model selection for LLM applications. Activate when the user needs to choose between different AI models, implement cost-efficient model routing, balance quality vs cost, or build intelligent model selection systems.

Use this skill when implementing caching for LLM applications. Activate when the user wants to reduce API costs through caching, implement semantic caching, cache LLM responses, optimize repeated prompts, or set up efficient caching strategies for AI applications.

Use this skill when securing AI applications against prompt injection. Activate when the user needs to prevent prompt injection attacks, validate AI inputs, implement input sanitization, or protect against adversarial prompts.

Use this skill when analyzing and optimizing LLM API costs. Activate when the user wants to reduce AI API spending, understand token usage, audit LLM costs, optimize prompts for cost efficiency, or track and report on AI expenditure.

Structure SKILL.md content so the model reads just enough — concise summary up front, progressively deeper detail, examples on demand. Covers section ordering, length budgets, when to split into multiple skills. Use this skill when writing or refactoring a skill body, one skill has grown too long, or a skill is wordy but not useful. Activate when: SKILL.md structure, skill content, skill too long, split skill, progressive disclosure, skill body.

Design skills that fire at the right moment — neither over-eager (noise) nor under-eager (silent). Covers activation specificity, trigger phrases, disambiguation between overlapping skills, and debugging activation. Use this skill when multiple skills could fire on the same query, a skill never fires, or a skill fires too often. Activate when: skill won't activate, skill over-activates, overlapping skills, skill triggers, skill selection, skill disambiguation.

Write the YAML frontmatter for a SKILL.md file so it activates reliably — name, description, and activation keywords that the model matches against. Covers length, tone, and the most common frontmatter mistakes. Use this skill when authoring a new skill, fixing a skill that isn't auto-activating, or reviewing skills for publication. Activate when: SKILL.md frontmatter, skill description, skill activation, skill YAML, write a skill, author a skill.

Test skills for correct activation, content quality, and regression — both automated checks (frontmatter validity, lint) and manual verification (query-suite activation testing). Covers CI integration and how to catch skill regressions before users do. Use this skill when adding skills to a repo, setting up CI for a skill library, or debugging "the skill exists but doesn't work". Activate when: test skills, validate skills, skill CI, skill linting, skill activation test, skill regression.

Implement checkpointing for agent recovery, debugging, and replay. Use this skill when building recoverable agents, implementing replay, debugging agent failures, or creating resumable workflows. Activate when: agent checkpoint, agent recovery, resume agent, agent restart, workflow replay, agent debugging, failure recovery, state snapshot.

Implement persistent agent state that survives failures and restarts. Use this skill when building stateful agents, implementing checkpointing, persisting agent memory across sessions, or recovering from failures. Activate when: durable state, agent persistence, checkpointing, agent recovery, stateful agents, state persistence, cross-session memory, agent restart.

Build agent workflows with LangGraph 1.0 state machines and graph patterns. Use this skill when creating agent graphs, implementing state machines, building multi-step agent workflows, or using LangGraph. Activate when: LangGraph, agent graph, state graph, agent workflow, graph nodes, conditional edges, agent state machine, ReAct agent.

Implement Agent-to-Agent (A2A) communication for cross-framework interoperability. Use this skill when building multi-agent communication, implementing agent protocols, connecting agents across frameworks, or standardizing agent interfaces. Activate when: agent to agent, A2A, agent communication, agent protocol, cross-framework agents, agent interoperability, MCP, agent discovery.

Use this skill when managing AI agent costs. Activate when the user needs to control token usage, implement cost limits for agents, optimize LLM spending, track agent costs, or prevent runaway API bills in agent systems.

Use this skill when implementing error handling for AI agents. Activate when the user needs agents to handle failures gracefully, implement retry strategies, design fault-tolerant agent systems, or build agents that can recover from errors without human intervention.

Use this skill when designing task handoffs between agents. Activate when the user needs to pass work between agents, transfer context between agents, implement agent-to-agent communication, or design protocols for agents to collaborate on sequential tasks.