Latest Agent Skills

Stand up a self-hostable, typo-tolerant search environment with Typesense — the open-source Algolia / ElasticSearch alternative (single C++ binary, <50ms instant search, no runtime deps). One routing-first skill: pick a server mode (binary download, official Docker image, or managed Typesense Cloud), install an API client (Python/JS/PHP/Ruby official; Go/Dart/C# community), design a collection schema, index documents, and run searches with typo tolerance, faceting/filtering, geo-search, sorting, grouping, synonyms, curation, scoped API keys, and federated multi-search — then wire an InstantSearch.js UI and a Raft-based HA cluster for production. Use when the user wants to build or operate an installable search backend, add site/app/product search, or migrate off Algolia/Elasticsearch. Triggers on: typesense, search engine, typo-tolerant search, algolia alternative, elasticsearch alternative, instantsearch, faceted search, geo search, vector search, self-hosted search, site search, product search.

Convert plans/specs into independently-grabbable vertical slice issues (HITL or AFK classification)

Route web-scraping work into the lightest workable Scrapling mode instead of defaulting to a browser. Use when the user needs HTML extraction, JS-rendered page retrieval, protected-target escalation, quick CLI scraping, agent-facing MCP access, or a larger crawl with Scrapling spiders. Triggers on: scrapling, scrape website, crawl site, adaptive scraping, selector drift, stealthy fetch, browser scraping, scrape to markdown, scrapling mcp, scrapling spider, research harvesting, literature scraping, paper metadata.

Install and route through OpenSpace, the skill management layer for AI agents, so a host agent can retrieve/rank/load the right SKILL.md out of this jeo-skills catalog (~150 installed skills), then evaluate skill quality from real execution evidence and evolve skills via FIX/DERIVED/CAPTURED updates. Covers install-as-skill-finder, retrieve-a-skill, evaluate-quality, evolve-skills, and local-first hub share/import. Triggers on: openspace, skill finder, skill retrieval, find the right skill, rank skills, skill discovery, skill quality, evolve skill, FIX DERIVED CAPTURED, skill hub, openspace-mcp, DiscoverSkills, skill-discovery, delegate-task.

Route LangSmith work into one workflow packet before touching SDK code. Use when the user needs LangSmith tracing, offline evals, annotation/review queues, prompt-registry decisions, audit/gap review, or cross-service trace propagation for an LLM app or agent workflow. Choose one packet: trace-debug, eval, review, prompt-registry, propagation, or audit. Triggers on: LangSmith, LangChain tracing, `@traceable` / `traceable`, `wrap_openai` / `wrapOpenAI`, datasets, experiments, annotation queues, feedback criteria, Prompt Hub, run trees, trace IDs, or production confidence for an AI feature. Not for generic SLO/alert design, non-LangSmith deployment orchestration, or runtime guardrails outside LangSmith.

Browse, group, relate, and selectively install the jeo-skills catalog through the lightweight `jeo-skill` CLI. Use when the user wants skills organized by web, infrastructure, game, creative media, CLI tools, AI/agents, engineering, research, business, or utilities; needs a frontend/backend/game-audio/game-VFX subcategory; wants overlapping skills connected instead of duplicated; or wants a category, bundle, or named skills installed without copying the full repository.

Cross-runtime 2-stage pipeline for Claude Code, Codex/OMX, and Gemini/Antigravity/OMA: trace causal hypotheses, inject evidence into deep-interview style requirements crystallization, then hand off to the right runtime planner/executor.

Run a reproduce → isolate → verify debugging workflow for concrete bugs, regressions, flaky failures, and environment-specific behavior. Use when the user already has a failing command, test, request, UI flow, or narrowed symptom and needs root-cause diagnosis or fix verification rather than raw log-line selection, broad test-policy design, PR review, or generic performance tuning.

Free offline desktop video/image compression (Tauri+React) — batch compress, trim/split, convert formats, embed subtitles. Install: brew install --cask codeforreal1/tap/compresso

Claude Code hook library — pre-built PreToolUse/PostToolUse hooks for common guardrails, auto-formatting, workflow automation

Turn raw ideas into one clear pre-planning concept artifact by choosing the right framing mode: problem framing, audience and value framing, concept shaping, game concept framing, or story packaging. Use when the user needs the next artifact before PRDs, sprint plans, launch execution, or game-production routing.

Get higher-level architectural perspective: maps modules, callers, dependencies using domain vocabulary

Route web-game auto-playtesting with WAI Play (waiterve/wai-play): decide whether the next move is a testability check, authoring or repairing the `GameFlowAgentAPI` bridge, running a real browser playtest, reading the five-dimension quality report, or unblocking a key node the agent cannot reach. Use when the user wants an AI agent to actually play their HTML5 / canvas / vibe-coded web game and return reproducible evidence, scores, and fix suggestions across the five supported types (survivor-like, arcade shooter, platformer, puzzle/card, visual novel). Triggers on: wai-play, WAI Play, auto-playtest, AI plays my game, web game testing agent, GameFlowAgentAPI, GameFlowIntegration, jumpToScenario, game quality score, playtest evidence. Route Unity/Unreal frame-time work to `game-performance-profiler`, engine build failures to `game-build-log-triage`, human playtest notes to `game-demo-feedback-triage`, and generic browser automation to `browser-harness`.

Run GitHub's Spec-Driven Development (SDD) workflow via the `specify` CLI — install spec-kit, initialize a project for one of 30+ AI coding agents (Claude Code, Copilot, Gemini, Cursor, Codex, Qwen, opencode, Kiro, etc.), and drive the constitution → specify → plan → tasks → implement command pipeline. Use when the user wants to bootstrap a Spec-Driven Development project, install `specify-cli`, generate executable specs before code, or invoke the `/speckit.*` slash commands (`/speckit.constitution`, `/speckit.specify`, `/speckit.plan`, `/speckit.tasks`, `/speckit.implement`, `/speckit.clarify`, `/speckit.analyze`, `/speckit.checklist`). Triggers on: spec-kit, speckit, specify, specify init, spec-driven, spec driven development, SDD, /speckit, executable spec.

Make the agent solve coding tasks with the least code that remains correct. Before writing code, walk the Ponytail ladder: skip what need not exist, then prefer stdlib, native platform features, already-installed dependencies, one line, and only then the minimum custom code. Use when the user asks for ponytail mode, less code, YAGNI, anti-bloat, minimal code, an over-engineering review, a current-diff delete-list, a whole-repo bloat audit, or a `ponytail:` tech-debt harvest. Keep validation, data-loss handling, security, and accessibility. Mark shortcuts with `ponytail:` plus the upgrade path. Triggers on: ponytail, /ponytail, /ponytail-review, /ponytail-audit, /ponytail-debt, write less code, YAGNI, over-engineering, anti-bloat, minimal code, do I need this, lazy dev.

Build, test, run, and flash NightRun — a bare-metal, no_std Rust UEFI application that boots straight into a local LLM (Llama 3.2, Qwen3, or Granite 4.1) with no operating system underneath. Use when the user wants to build/flash a bootable NightRun USB or Raspberry Pi 5 SD image, convert a GGUF model into the `.nrm` container, run/debug the inference engine on the host or in QEMU/OVMF, or troubleshoot no_std kernel/tokenizer parity issues in the NightRun codebase. Triggers on: "nightrun", "boot into an LLM", "bare-metal LLM runtime", "UEFI LLM appliance", "nrconvert", "nrhost", "cargo xtask", "nrm model file", "flash a bootable LLM USB".

Build and maintain a persistent markdown wiki that an LLM updates on the user's behalf, usually inside an Obsidian vault or git-tracked notes repo. Use when raw sources such as web articles, papers, meeting notes, transcripts, screenshots, or past analyses need to be turned into an interlinked knowledge base with immutable source files, LLM-written wiki pages, `index.md`, `log.md`, schema rules in `AGENTS.md` or `CLAUDE.md`, source summaries, query notes, and recurring lint passes. Triggers on: llm-wiki, personal wiki, obsidian wiki, research vault, knowledge base, source ingest, persistent notes, wiki maintenance, source summaries, query filing.

Systematic plan stress-testing through relentless one-question-at-a-time decision-tree interviewing

Drive Graphify from its CLI to build, refresh, query, export, and serve a durable code/corpus knowledge graph. Use when the user wants `.graphify/GRAPH_REPORT.md`, `graph.json`, `graph.html`, `graphify update`/`summary`/`query`/`path`/`explain`/`tree`, change-aware review context, git-hook or watch-based refresh, a stdio MCP graph server, or an install into jeo, jeopi, gjc, opencode, Claude, Codex, or Gemini. Also covers the honest structural fallback when native extraction is empty or misleading. Route simple locate/reference work to `codebase-search`, narrative knowledge-base work to `llm-wiki`, and project-memory handoff to `opencontext`. Triggers on: graphify, graphify update, graphify query, knowledge graph CLI, GRAPH_REPORT.md, graph.json, codebase graph, graph refresh, graphify install, graphify serve, review context, affected flows.

Route goalflow (wanmol/goal-flow) work — a LangGraph framework that combines workflow graphs with agent loops — into exactly one mode: fit check, transpiling a Dify DSL export into runnable LangGraph Python, authoring workflow nodes and edges, building an `agent_kit` loop with middleware and a harness, wiring the serving layer (data adapters, SSE streaming, HITL, Redis/MySQL, API-key registration), or running the pre-publish security gate. Use when the user wants Dify's visual design without Dify's runtime, a graph node that hosts an agent loop, an OpenAI-compatible wire protocol over their own workflows, or prompt-injected `SKILL.md` capabilities. Triggers on: goalflow, goal-flow, dify to langgraph, dify transpiler, dify DSL export, BaseWorkflow, agent_kit, AgentBaseNode, DataAdapter, chunk processor, HITL interrupt, dify2langgraph. Route plain graph-API questions to `langgraph-fundamentals` and `langgraph-workflow`.

Design, implement, budget, and verify readable real-time game visual effects: particles, smoke/fog, lightning, fire, trails, impacts, bloom/glow, and layered spell timelines. Use when a user needs a VFX spec, lifecycle, engine handoff, particle pooling, effect composition, reduce-motion fallback, or frame-budget diagnosis. Triggers on: game VFX, spell effects, particle system, impact burst, procedural lightning, smoke, bloom, magic aura, or VFX performance.

Investigates hard, flaky, or performance failures through a six-phase feedback-loop-first debugging process; use when quick targeted debugging is insufficient.

[devup-ui] Routing-first front door for Devup UI — a zero-runtime CSS-in-JS library whose Rust + WebAssembly preprocessor extracts every style at build time, so no styling JavaScript ships at runtime. Use to adopt/integrate Devup UI, wire the build-time plugin into Next.js/Vite/Rsbuild/Webpack/Bun, write styles with Box/css props or the styled-components-compatible styled() API, set up type-safe theming with devup.json, or migrate off styled-components/Emotion/Tailwind/Panda/vanilla-extract to zero runtime.

Create or refresh hierarchical AGENTS.md documentation for Claude Code, Codex/OMX, Gemini, and Antigravity/OMA projects, preserving manual notes while excluding runtime state such as root .omc, .omx, .survey, .codex, and generated build folders.

Turn any GitHub repo, local folder, PR, or markdown/Obsidian vault into an interactive architecture map with CodeFlow — a zero-build single index.html browser app (React 18 + D3.js from pinned CDNs) that runs 100% client-side with no backend and no data collection. Pick an input (public repo, private repo with a local token, local files, PR URL, markdown vault), choose a visualization mode (folder/layer/churn/blast), and read the analysis — dependency graph, blast radius, code ownership, heuristic security scanner, pattern/anti-pattern detection, A–F health score, activity heatmap, PR impact — then export JSON/Markdown/text/SVG/PDF or wire the self-updating CodeFlow Card SVG onto a README.

Drive CodeBurn, a free open-source local-first CLI/TUI/web/menubar tool that reads the session files already on disk from 40 AI coding tools (Claude Code, Codex, Cursor, Gemini CLI, Grok, OpenCode, and more) and breaks down token usage and dollar cost by task, model, tool, and project. Use when the user wants to see where their AI coding spend went, find and fix token waste in a Claude Code / agent setup, cap a session's budget before it runs away, compare which model is actually worth its price, check whether AI spend shipped or was reverted, or wire live usage/savings data into an agent over MCP. Triggers on: "codeburn", "npx codeburn", "AI token usage", "AI coding cost", "where did my Claude spend go", "codeburn optimize", "codeburn guard", "codeburn compare models", "codeburn yield", "token waste in CLAUDE.md", "AI spend dashboard".

AI-driven Game Development Studio (BMAD-GDS). Routes game projects through Pre-production, Design, Architecture, Production, and Game Testing phases using 6 specialized agents. Supports Unity, Unreal Engine, Godot, and custom engines.

Self-hosted AI gateway with one OpenAI-compatible endpoint for multi-provider LLM, embedding, image, and audio routing, automatic fallback, load balancing, and cost optimization.

Turn a weak/cheap "Flash" model into a "Pro" performer by wrapping HKUDS UpSkill — captures agent session failures, has a strong Teacher model analyze them and draft a skill, then validates it against the weak Student model in a closed Ralph Loop (up to 3 rounds) before storing it for automatic reuse. Use when the user wants to install UpSkill, run `/upskill-init`, `/upskill-configure`, `/upskill-build`, `/upskill-run`, `/upskill-list`, `/upskill-status`, `/upskill-mode`, `/upskill-model`, `/upskill-remove`, or `/upskill-uninstall`, wants a cheap model to perform closer to a Pro model without switching, or wants a good session (success or failure) distilled into a validated skill. Triggers on: upskill, up-skill, flash to pro, teacher student distillation, ralph loop skill validation, distill agent failures into skills. Routes skill-quality ratcheting to `skill-autoresearch`, scaffolding to `write-a-skill`, and spec-compliance rewrites to `skill-standardization`.

Fast, accurate, and comprehensive multiline text measurement and layout library for JavaScript/TypeScript — without DOM reflow. Use when the user needs to calculate paragraph heights, line counts, manual line layouts (text flowing around floated images), or rich text rendering (emoji, CJK, RTL) on DOM, Canvas, or SVG targets. Covers npm install, core API (prepare/layout, prepareWithSegments/layoutWithLines), and options (whiteSpace, wordBreak, letterSpacing). Triggers on: pretext, text measurement, text layout, paragraph height, line layout, text reflow, canvas text measurement, DOM reflow.

Work with the OpenCut open-source video editor repo (github.com/OpenCut-app/OpenCut) — clone/setup, run the web/desktop dev servers, understand the Rust/WASM compositor core, and follow current contribution focus areas. Use when the user asks about OpenCut, the "open-source CapCut alternative", opencut-classic, setting up the OpenCut monorepo (moon/proto or Bun/Docker), the OpenCut Rust core, or contributing a PR to OpenCut. Also triggers on "opencut.app", "new.opencut.app", and "opencut-app/opencut-classic".

Lazyweb is the design-evidence skill for AI coding agents. Use it when designing, critiquing, or changing product UI — it provides real app screenshots, competitor references, best practices, quick examples, creative cross-category ideas, paywall optimization guidance, and mobile growth + monetization A/B test context. Routes to the right Lazyweb mode (design workflow, quick search, update, flowchart, or A/B test research) using MCP tools and the Lazyweb server at https://www.lazyweb.com/mcp. Use before designing any screen when you need design evidence instead of training-data vibes.

Runs directional ablation and refusal-direction analysis for open-weight models the user may modify; use to reduce benign over-refusal or measure refusal/KL trade-offs, not for training.

Design domain-specific agent teams, define specialized agents, and generate the skills they use. Use when you need to decompose a complex project into coordinated multi-agent teams, choose the right architecture pattern (pipeline, fan-out/fan-in, expert pool, producer-reviewer, supervisor, hierarchical delegation), generate .claude/agents/ and .claude/skills/ files, or validate and iterate on generated harnesses. Triggers on: harness, build a harness, design agent team, agent team architecture, multi-agent skill generation, set up harness, harness engineering, domain agent team, harness for this project.

Assist with Colibri: pure-C LLM inference engine for running GLM-5.2 (744B MoE) on consumer machines with ~25 GB RAM. Use when setting up, building, converting models, running inference, configuring expert streaming and caching, optimizing speculative decoding (MTP), GPU integration, and integrating Colibri into production pipelines. Includes build setup, model download & conversion, chat/inference modes, performance tuning, and API integration patterns.

Enables persistent ultra-compressed technical communication; use for explicit brevity or token-reduction requests, except where fragments risk safety or clarity.

Drive Animato (github.com/otdnnc/Animato) as an API-key agent loop that turns a rigged .fbx/.gltf model plus a plain-text motion request into a baked animation: upload the model, build the bpy prompt, spend one LLM call with your own key, gate the generated script, run it headless, and verify the animated output. Use when the user wants text-to-animation for a 3D character, an unattended animation pipeline driven by a Gemini or OpenAI-compatible API key, or help operating a local Animato server. Triggers on: animato, text to animation, animate a rigged model, bpy animation script, blender headless keyframe, /api/chat animation, character motion from a prompt, GEMINI_API_KEY animation.

Drive Airship (`@airshiplabs/cli`), a CLI that puts a visual, pannable design canvas in front of a running dev server so an engineer can click an element, describe a change, and have Claude Code, Codex, or OpenCode edit the underlying source file directly — no plugin, no build config, nothing added to the project. Use when the user wants to launch a visual editor over a local dev server (`npx @airshiplabs/cli --target <port>`), pick or compare coding agents (`--agent claude|codex|opencode`) for live UI edits, sandbox agent edits with `--safe`, scaffold an `airship.config.json` (`airship init`), or diagnose a broken setup (`airship doctor`). Triggers on: "airship", "@airshiplabs/cli", "visual editor for my codebase", "click to edit UI", "design canvas over dev server", "airship doctor", "airship init", "pair Claude Code with a visual editor on localhost".

Automatically rig 3D models with UniRig (VAST-AI-Research, SIGGRAPH'25) — predict a skeleton, predict skinning weights, and merge the rig back onto the original mesh. Use when the user wants auto-rigging for .obj/.fbx/.glb/.gltf/.dae/.vrm assets, a skeleton or skin weights for a character or creature, a UniRig environment prepared on a CUDA machine, batch rigging of a model directory, or an honest comparison between UniRig, SkinTokens, Tripo, Mixamo, AccuRig, and Blender Rigify. Triggers on: unirig, auto rig, auto-rigging, 3D rigging, skeleton prediction, skinning weights, rig a character, armature generation, rigged glb, rigged fbx, bone weights.

Moves tracker issues through a structured triage state machine with category and readiness decisions; use to evaluate or prepare work items for implementation.

Set up, run, and contribute to TokHub (github.com/yaojingang/TokHub) — an open-source AI API relay monitoring, recommendation, and OpenAI-compatible gateway system with L1/L2/L3 channel health probing, usage metering, alerts, audit, and Docker self-hosting. Use when the user asks about TokHub, "AI API 中转站监控", cloning/running the Go + React monorepo (TOKHUB_ROLE, sqlc, TimescaleDB, NATS), the L1/L2/L3 probe algorithm, the OpenAI-compatible `/gateway/v1/*` endpoint, or contributing a PR to TokHub. Do not use for connecting a running agent to a live TokHub instance's own API (that is covered by the project's own bundled `agent-skills/tokhub` skill inside the TokHub repo, not this one).

Generate structured PRD from conversation context without interviewing the user

Run a bounded cross-platform landscape scan before planning or implementation. Use when the real job is researching what exists, how people work around it, which solutions repeat, or how platform/tooling patterns map before deciding what to build. Produce reusable `.survey/{slug}/` artifacts, validate the artifact contract, and route planning or execution outward only after the survey is done.

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.

Generate, convert, preview, and integrate retro game sound effects with raysan5/rfxgen. Use when a user needs coin, laser, explosion, power-up, hit, jump, or blip SFX; wants `.rfx` parameters converted to WAV/raw/C headers; needs batch CLI generation; or must build and troubleshoot rFXGen. Triggers on: rFXGen, rfxgen, sfxr, chiptune SFX, .rfx, procedural game sound, or sound preset.

Run KADATH (Kernel for Agentic Darwinian Adaptation, Tooling, and Heredity), a Docker-based evolutionary kernel that turns a goal into a locked, Architect-authored benchmark, then evolves a population of smolagents-based coding agents across epochs: each agent runs in an isolated container, gets graded against frozen evidence, and the population is culled, mutated, and reproduced generation over generation until it converges on the best-performing agent framework for that goal. Use when the user wants to propose/approve/run a KADATH evolutionary run, check a run's status or live dashboard, pause/resume/continue a run, export the winning agent population, or understand its Architect/Grader/Tweaker/Birther pipeline, evidence-freezing, or genome lineage/memory model. Triggers on: "kadath", "kadath.sh", "evolve an agent", "Darwinian agent evolution", "agent population fitness benchmark", "smolagents evolutionary run", "kadath dashboard", "genome lineage", "epoch champions".

Search and download specific files/folders from GitHub repositories directly from terminal using ghgrab, without full clone. Covers install, interactive browsing, release asset download, and automation-safe usage patterns.

Route Firebase AI feature work into either direct app/client Firebase AI Logic SDK integration or a server-owned Genkit workflow. Use when a web, mobile, backend, or full-stack feature needs model calls, typed outputs, reusable flows, tools, retrieval, prompt files, evals, observability, or deployment. Choose client-ai-logic, flow-foundation, tool-and-agent, retrieval-and-prompt, evaluation-and-observability, deployment-runtime, or comparison-or-fallback; route Firebase platform/operator work to `firebase-cli` and broad framework comparisons to `survey`.

Routing-first Fabric operator skill for reusable named AI transforms over stdin, files, transcripts, notes, logs, and cleaned web text. Use when the user wants Fabric patterns, custom pattern packs, shell-pipe composition, or `fabric --serve` workflows — not generic one-off chat prompting, repo-aware coding, or fully deterministic automation.

Turn natural-language descriptions into editable `.drawio` diagrams and export them to PNG / SVG / PDF / JPG via the native draw.io desktop CLI, or turn an existing codebase (Python / JS-TS / Go / Rust) into an auto-laid-out structure diagram. Wraps Agents365-ai/drawio-skill: 6 diagram presets (ERD, UML class, sequence, architecture, ML/DL, flowchart), search across 10,000+ official AWS/Azure/GCP/Cisco/K8s/UML/ BPMN shapes, 321 AI/LLM brand logos, vision self-check + auto-fix, and a 5-round iterative refinement loop. No MCP server, no daemon — runs from a single SKILL.md and the draw.io CLI. Use when the user wants polished, precise, exportable diagrams or wants to visualize code structure. Triggers on: drawio, draw.io, drawio diagram, architecture diagram, ERD, UML diagram, sequence diagram, flowchart, network diagram, visualize codebase, code structure diagram, class hierarchy, export diagram png/svg/pdf, AWS/Azure/GCP icon, draw.io shapes.

Build fully customizable, agent-ready design systems with Astryx — Meta's production design system now open source. Ships 150+ React components built on StyleX with zero styling lock-in, component swizzling, brand theming, dark mode, and CLI tooling. Use when building component libraries, design systems, UI applications, design tokens, or when teams need consistent accessible components that AI agents can understand and extend. Triggers on: astryx, design system, component library, design tokens, react components, accessible components, astryx design, stylesheets, theme customization, component composition.

Enterprise AI adoption platform for Claude Code — measures team usage via OpenTelemetry, syncs skills/MCP servers/hooks from a central dashboard, and delivers context-aware skill suggestions at prompt time. Solves the Intention-Action Gap: organizations that deploy Zeude see 3x adoption improvement (6% → 18%). Requires Supabase (config) and ClickHouse (analytics). Triggers on: zeude, ai adoption, claude code adoption, enterprise claude, opentelemetry claude, skill sync, team claude management.

Install, configure, and operate Strix for AI-driven application security testing. Use when you need to run authorized vulnerability scans against local codebases, GitHub repositories, staging URLs, domains, or CI pipelines; configure Docker and LLM providers; choose quick, standard, or deep scan depth; or pass authenticated testing instructions to Strix. Triggers on: strix, ai pentest, vulnerability scan cli, appsec scan, bug bounty automation, strix ci, strix docker, strix scan mode, strix instruction file, headless security scan.

Drive Soup (`soup-cli`), a CLI-first tool for fine-tuning and post-training LLMs with one YAML config and one command — SFT, DPO/GRPO/ORPO/SimPO/KTO, QLoRA/DoRA/LoRA+, layer streaming for 4-8 GB GPUs, eval-gated training, and serving. Use when the user wants to `soup init`/`soup train` a model, pick a training method or quantization scheme, estimate cost/memory before training, fine-tune on a small local GPU, migrate a config from Axolotl/LLaMA-Factory/Unsloth, or serve/merge/push a trained adapter. Triggers on: "soup-cli", "soup train", "soup init", "fine-tune an LLM locally", "QLoRA on a laptop GPU", "layer streaming", "soup advise", "soup autopilot", "DPO/GRPO/ORPO training", "merge LoRA adapter".

Run Comet's Opik — open-source LLM observability, evaluation, and optimization — from one routing-first skill: install the Python/TypeScript SDK, stand up a server (Comet.com cloud, Docker Compose via `./opik.sh`, or Kubernetes/Helm), wire tracing through `@opik.track` or one of 50+ framework integrations (OpenAI, Anthropic, LangChain, LangGraph, LlamaIndex, CrewAI, DSPy, Ollama, Bedrock, Vercel AI SDK, …), score outputs with LLM-as-a-judge metrics (Hallucination, Moderation, Answer Relevance, Context Precision), and run Datasets/Experiments evaluations including PyTest CI gates. Use when the user wants LLM tracing, prompt evaluation, production LLM monitoring, agent optimization, or guardrails with Opik. Triggers on: opik, comet opik, opik configure, opik.sh, llm observability, llm tracing, llm as a judge, hallucination metric, prompt evaluation, opik dashboard, opik guardrails, agent optimizer.

Drive Moli (`moli`), Lexmount's open-source headless browser for AI agents, built around on-demand rendering: real JavaScript, DOM, and CSS by default, with layout and pixels computed only when explicitly requested via `--layout`. Use when the user wants to fetch/extract a live JavaScript-rendered page as Markdown/HTML/JSON/semantic-tree, capture a screenshot or PDF, run a small bounded crawl, start a CDP/WebDriver automation server for Playwright/Puppeteer, replace a Chromium/ChromeDriver dependency, or diagnose readiness/network/frame issues on a rendered page. Triggers on: "moli fetch", "moli serve", "headless browser for agents", "on-demand rendering browser", "CDP server without Chrome", "structure-first web scraping", "Lexmount browser", "moli-webfetch", "moli-cdp-server".

Packet-first BMAD/BMM front door for idea notes, product briefs, PRDs, architecture drafts, review feedback, existing repo state, and milestone pressure. Use when the user wants to know what BMAD phase or artifact comes next, or needs a portable BMAD entrypoint before routing review, execution slicing, runtime setup, or game-production work outward.

Red-green-refactor TDD using vertical slices — tests specify observable behavior through public interfaces

Install, initialize, verify, and troubleshoot RTK (Rust Token Killer) for AI coding agents. Use when you need to reduce shell-command token output, confirm that the correct `rtk` binary is installed, choose between Homebrew, install.sh, or Cargo installation, wire `rtk init` for Claude Code, Codex, Gemini CLI, Cursor, Copilot, Windsurf, Cline, or OpenCode, or use compact wrappers such as `rtk git status`, `rtk read`, `rtk grep`, `rtk test`, `rtk lint`, and `rtk gain`. Triggers on: rtk, rust token killer, token saver cli, rtk init, rtk gain, codex rtk, gemini rtk, opencode rtk, claude hook token reduction.

Run the Ouroboros specification-first development loop: reduce ambiguity with a Socratic interview grounded in live git data (commits, churn, contributors), freeze an immutable seed/spec, render the execution plan through spec-kit (/speckit.plan → /speckit.tasks), execute against that contract through cli-anything agent-native CLI harnesses (cli-hub, --json evidence), verify before claiming success, and keep looping until completion is actually verified. Use when the user wants spec-first clarification, git-aware interviews, immutable requirements, drift-aware implementation, harness-driven execution, or a persistent completion loop that should keep going until tests / checks / acceptance criteria pass. Triggers on: ooo, ouroboros, interview, seed, run workflow, evaluate, evolve, ooo ralph, specification first, socratic interview, git-aware interview, ambiguity reduction, execution plan, cli harness execute, persistent completion.