skills

Expert guidance for implementing the A2UI (Agent to UI) protocol — covering surfaces, components, data binding, catalogs, message streams (v0.9 / v0.9.1: createSurface / updateComponents / updateDataModel / deleteSurface; v1.0 adds callFunction / functionResponse / actionResponse; v0.8: surfaceUpdate / dataModelUpdate / beginRendering), the A2A extension binding, custom component catalogs (schemas, renderers, negotiation, versioning, graceful degradation), client/renderer architecture (MessageProcessor / SurfaceModel / ComponentImplementation / Binder), custom functions, the sendDataModel flag, surfaceProperties, and v0.8 → v0.9 → v0.9.1 → v1.0 migration. Always invoke for any question that mentions A2UI, surfaceUpdate, beginRendering, createSurface, updateComponents, updateDataModel, sendDataModel, callFunction, functionResponse, actionResponse, surfaceProperties, a2uiClientCapabilities, supportedCatalogIds, catalogId, basic_catalog, formatString, ChildList, ComponentId, or agent-driven UI streaming.

Expert guidance for the A2A (Agent2Agent) Protocol — an open HTTP-based protocol for independent, opaque AI agents to collaborate as peers. Covers v1.0: data model (Agent Card, Task, Message, Part, Artifact, Extension); operations (`SendMessage`, `SendStreamingMessage`, `GetTask`, `ListTasks`, `CancelTask`, `SubscribeToTask`, the `*TaskPushNotificationConfig` family, `GetExtendedAgentCard`); task lifecycle and multi-turn flows; three protocol bindings (JSON-RPC, gRPC, HTTP+JSON/REST) and custom bindings; agent discovery (well-known URI, registries, signed Agent Cards); enterprise security (TLS, OAuth 2.0 / OIDC / mTLS, in-task authorization); SSE streaming and webhook push notifications; the extension framework; A2A↔MCP relationship; and the six SDKs (`a2a-python`, `a2a-js`, `a2a-java`, `a2a-go`, `a2a-dotnet`, `a2a-rs`). Always invoke for A2A, Agent2Agent, agent card, A2A tasks / messages / artifacts / parts, push notification configs, A2A extensions, Agent Card signing, or any of the six SDK package names.

Expert guidance for AG-UI (Agent–User Interaction Protocol) — the open, event-based protocol connecting AI agents to user-facing apps. Covers the run lifecycle, event types (lifecycle, text, tool-call, state, activity, reasoning), RunAgentInput, AbstractAgent/HttpAgent, multimodal messages, shared state via STATE_SNAPSHOT/STATE_DELTA (RFC 6902 JSON Patch), frontend-defined tools, interrupt-aware HITL with RunAgentInput.resume, encrypted reasoning for ZDR, capability discovery, middleware, serialization with parentRunId branching, the TypeScript (`@ag-ui/client`, `@ag-ui/core`) and Python (`ag-ui-protocol`) SDKs plus community SDKs. Use whenever someone builds or consumes AG-UI, wires up CopilotKit, LangGraph, CrewAI, Mastra, Pydantic AI, LlamaIndex, Agno, AG2, Microsoft Agent Framework, Google ADK, AWS Strands, or Bedrock AgentCore, or mentions AG-UI, ag_ui, RunStartedEvent, STATE_DELTA, TOOL_CALL_START, REASONING_*, AbstractAgent, HttpAgent, or EventEncoder. Always invoke.

Expert guidance for the Model Context Protocol (MCP), the JSON-RPC 2.0 protocol connecting LLM apps to external tools and data. Covers spec versions 2024-11-05 through 2025-11-25 (architecture, transports, Resources, Prompts, Tools, Sampling, Roots, Elicitation, Tasks, lifecycle, OAuth 2.1, utilities) plus the official language SDKs: the Rust SDK `rmcp` (full guide — `ServerHandler`/`ClientHandler`, `#[tool_router]`/`#[tool_handler]`/`#[task_handler]` macros, `StreamableHttpService`, Cargo features, SEP-1319 tasks, `tokio::io::duplex` test harness) and the TypeScript SDK `@modelcontextprotocol/sdk` (incl. the `pkce-challenge` Vite/bundler resolver failure). Use when building MCP servers or clients in any language, working with transports or OAuth, or debugging SDK-specific issues. Always invoke for questions mentioning MCP, modelcontextprotocol, `rmcp`, tools/list, tools/call, sampling/createMessage, elicitation/create, Streamable HTTP, Mcp-Session-Id, `pkce-challenge`, or `@modelcontextprotocol/sdk`.

Eval-driven prompt refinement that runs entirely inside Claude Code via the Agent/Task tool — no Python, no SDK, no API key. Each candidate run and each judge call executes in an isolated subagent with a fresh context window, so samples are independently graded and the main session stays focused on synthesis and iteration. Trivially parallel: spawn N candidate + M judge subagents in one assistant message. Invoke when the user wants to evaluate, A/B test, regress-test, or iterate on a prompt directly inside Claude Code, especially when they reference subagents, the Task/Agent tool, or "test this prompt without writing code". Phrases like "use Claude Code to evaluate this prompt", "spawn subagents to test", "parallel-test these variants", "A/B these prompts in Claude Code", "grade this rubric with subagents", or "iterate this prompt with fresh contexts" qualify. Pairs with the broader `prompt-evaluation` skill for shared dataset-design and binary-judge methodology.

Org-internal engineering standards for this organization's projects. Invoke whenever scaffolding a new repo, auditing an existing one, setting up CI, or configuring tooling files like `Cargo.toml`, `rust-toolchain.toml`, `justfile`, `.editorconfig`, `.markdownlint-cli2.yaml`, `tsconfig.json`, `package.json`, `pnpm-lock.yaml`, `bunfig.toml`, `pyproject.toml`, `uv.lock`, `.python-version`, or `*.tf`. Also invoke for work involving `axum`, `utoipa`, `markdownlint-cli2`, `uv`, `ruff`, `pyright`, `pytest`, or Node package-manager setup (pnpm is the org default). Rust and Python are fully documented: Cargo workspace inheritance, MSRV pinning, `just` as task runner, `cargo-llvm-cov` coverage, hermetic-vs-live test split, Axum + utoipa OpenAPI; uv workspaces, packaged `src` layout, ruff, pyright strict, pytest live-marker tiers. TypeScript, Node.js, Bun, Terraform, Rust libraries, and Rust GraphQL are stubs — invoke anyway so the user can define the convention rather than receive an improvised one.