Review a Confluent agent skill in this repo against the Agent Skills spec (agentskills.io), Confluent conventions in CLAUDE.md, the PR template gates, and the evals-as-contract rule. Use this skill whenever the user asks to review, audit, validate, or lint a skill; opens or inspects a PR that adds or modifies anything under `skills/`; asks about spec conformance, lazy-loading, frontmatter shape, trigger overlap, or eval coverage; or wants a pre-merge sanity check on skill changes. Do NOT trigger for general code review of application code; security review; auditing schemas, producer/consumer configs, PII tagging, or Terraform generation for Schema Registry (handled by `kafka-schema-registry`); runtime/log analysis of skill behavior (use `tools/skill_review_dashboard.py`); or any changes that don't touch the `skills/` tree.

Reviews and improves Claude Code skills against official best practices. Supports three modes - self-review (validate your own skills), external review (evaluate others' skills), and auto-PR (fork, improve, submit). Use when checking skill quality, reviewing skill repositories, or contributing improvements to open-source skills.

libllm - LLM API client for OpenAI-compatible endpoints. LlmApi class handles chat completions and embeddings via HTTP. Supports GitHub Models, Azure OpenAI, and standard OpenAI endpoints. Handles streaming responses, token counting, and multi-tool parallel call fixes. Use for LLM completions, embeddings, and AI model integration.

Use when the user has an idea for an app, tool, or feature and wants to be interviewed in depth first — answering structured clarifying questions until the requirements are fully pinned down — before it gets built end-to-end. Triggers - "ask me questions then build", "interview me about my idea", "help me spec this out and build it", vague ideas the user wants to shape interactively.

Use for Vercel cost and performance optimization on deployed projects, especially Next.js, SvelteKit, Nuxt, and limited Astro apps. Collect Vercel metrics, usage, project config, and code scan results first; investigate only metric-backed candidates; produce ranked recommendations grounded in verified files and version-aware Vercel/framework docs. Trigger for Vercel bill reduction, slow or expensive routes, caching opportunities, Function Invocations, Build Minutes, Fast Data Transfer, Core Web Vitals, Bot Management, Fluid compute, or cost breakdown requests.

Use when someone wants to find a recorded Paper session, theirs or anyone's in the org — "did anyone work on X", "find the session where…", "search paper sessions", "which session was that", "pull up that agent run" — or to browse sessions by time, author, or harness when there is no phrase to search. Covers `paperctl search` and `paperctl sessions` over recorded sessions (tapes).

Utility tools via uni CLI. Use when user wants weather, stock prices, currency conversion, QR codes, or URL shortening. Services: weather, stocks, currency, qrcode, shorturl.

uni CLI overview and meta commands. Use for general uni questions, multi-command execution (uni run), saved flows, natural language (uni ask), plugins, config, aliases, history. For specific services, use their dedicated skills (uni-wa, uni-telegram, uni-google, etc.)

Use when parallelizing development, running multiple agents, splitting work across agents, coordinating parallel tasks, or decomposing PRDs for concurrent execution. Breaks work into independent agent workstreams.

Reference for configuring tool permissions when launching Claude Code agents. Use when setting up --allowedTools flags, restricting file access, or configuring agent permissions.

Inspect a Kotlin + Spring repository before making changes or answering implementation, debugging, build, architecture, or review questions. Use when the agent must map modules, Spring Boot/Kotlin/Gradle/JDK versions, compiler plugins, profiles, dependencies, runtime stack, and architectural boundaries so later advice stays compatible with the real project.

Design resilient HTTP, messaging, and scheduled integrations for Kotlin plus Spring services with explicit timeout budgets, retries, idempotency, circuit breakers, DLQ behavior, and failure observability. Use when integrating with unstable external systems, designing retry logic, handling duplicate delivery, preventing thundering herds, or making message-driven and scheduled workflows safe in production.

Decompose business scope into bounded contexts, module or service boundaries, workflows, and API contracts before implementation begins. Use when shaping a new feature, service, or major redesign; when requirements are underspecified; when consistency, idempotency, and integration boundaries must be made explicit; or when ADR-quality tradeoff reasoning is needed for Kotlin plus Spring systems.

Investigate `pnpm test` memory growth, Vitest worker OOMs, and suspicious RSS increases in OpenClaw using the `scripts/test-parallel.mjs` heap snapshot tooling. Use when Codex needs to reproduce test-lane memory growth, collect repeated `.heapsnapshot` files, compare snapshots from the same worker PID, distinguish transformed-module retention from real data leaks, and fix or reduce the impact by patching cleanup logic or isolating hotspot tests.

Generate one or more standalone Meta image-ad creatives via Nano Banana 2 / Nano Banana Pro (Gemini Flash Image family) through the Arcads external API. Locks the model family, auto-strips platform chrome, enforces edge-safe layouts. Use when the user asks for a "Nano Banana ad", "Gemini image ad", "nano-banana-2 ad creative", "make a static image ad with Gemini", or anchors on a need for photoreal / lifestyle / multi-reference / handheld-object / clay-texture ad creatives (sticky-note flatlays, held-whiteboard signs, lifestyle portraits, ingredient collages, OOH photography). Does NOT trigger on ChatGPT Image cues — use chatgpt-image-ad for those.

Build presentation plans for PPT / slides / decks through a conversation-first workflow, then propose multiple visual directions with preview images before writing deck specs. Use when the user asks to create a PPT, presentation, deck, 答辩稿, 路演 deck, 产品介绍 PPT, 汇报 PPT, or when the user has only a topic or rough materials and needs help clarifying structure, style, and page planning before generation.

Redesign GitHub README homepages or create project-native pure SVG, hybrid SVG-composed PNG/WebP, and opt-in animated GIF assets. Use when a user asks to beautify, redesign, rebrand, visually upgrade, simplify, or audit a GitHub README; create only a hero, section headers, diagrams, badges, motion graphics, showcase modules, or other README assets; or turn a repository homepage into a cohesive visual story. If whole-README work versus asset-only work is unclear, ask which scope the user wants. For hero-like assets where pure SVG and generated raster material are both viable, explain the tradeoffs and confirm the implementation before creating the asset.

Convex backend development guidelines. Use when writing Convex functions, schemas, queries, mutations, actions, or any backend code in a Convex project. Triggers on tasks involving Convex database operations, real-time subscriptions, file storage, or serverless functions.

Transforms conversations and discussions into structured documentation pages in Notion. Captures insights, decisions, and knowledge from chat context, formats appropriately, and saves to wikis or databases with proper organization and linking for easy discovery.

Battle-tested PyTorch training recipes for all domains — LLMs, vision, diffusion, medical imaging, protein/drug discovery, spatial omics, genomics. Covers training loops, optimizer selection (AdamW, Muon), LR scheduling, mixed precision, debugging, and systematic experimentation. Use when training or fine-tuning neural networks, debugging loss spikes or OOM, choosing architectures, or optimizing GPU throughput.