skills

Produce Mermaid diagrams styled with my dark theme — a #171717 canvas, light hairline borders, and a fixed categorical palette via classDef. Use whenever I ask for a Mermaid diagram, flowchart, sequence/state/ER/class diagram, an architecture/system diagram, a README diagram, or say "mermaid", "dark theme diagram", "다이어그램 그려줘", "머메이드", "아키텍처 다이어그램". Default to this dark theme unless I explicitly ask for the default/light Mermaid look. Always emit the `%%{init}%%` header and reuse the canonical classDef library for diagram grammars that support it.

Use Grok Build's current web_search/web_fetch tools as a bounded external-evidence layer inside coding, debugging, dependency evaluation, and autoresearch loops. Use for "Grok으로 검색해", fresh release/API/CVE evidence, recent community signal, or search-grounded hypotheses. Treat dedicated X corpus search as optional and only claim it when an installed MCP/tool explicitly provides it; never leak private repo data.

Automate Grok Build with current headless (-p/--single), JSON/streaming output, resumable sessions, bounded toolsets, permission rules, worktrees, and ACP. Use for "grok -p", CI/cron/batch reviews, unattended refactors, Grok as an engine, or custom ACP clients. Verify flags against the installed Grok docs and prefer explicit tool allowlists plus deny rules over unrestricted execution.

Execute a repository milestone from GOAL.md through parallel implementation, staged review, CI, PR merge, and GitHub milestone closure. Use when Aaron asks to start, deliver, or complete the next milestone.

Keep unrestricted/YOLO Orca and coding-agent execution fast while constraining destructive filesystem actions, secret exposure, external writes, and confidential data disclosure. Use whenever an agent runs with bypassed approvals or sandboxing, before rm/cleanup commands, infrastructure or remote writes, secret-bearing commands, force-push/history rewrites, or work in confidential repositories.

Use when building, reviewing, debugging, validating, or safely deploying n8n workflows with the czlonkowski/n8n-mcp MCP server. Trigger on requests like "n8n workflow 만들어줘", "n8n-mcp 연결", "n8n 노드 찾아줘", "n8n 자동화 설계", "n8n credential/workflow 관리", or any task where an agent needs n8n node docs, templates, workflow JSON validation, or n8n API management. Always prefer read-only documentation/search first, work on copied workflows before production edits, and validate before deploy.

Orchestrate Mori-led multi-agent work across Hermes, Codex, Claude Code, Grok/headless, and agy for research/code/report workflows.

Analyze an open-source GitHub repository (such as CLI tools, Python scripts, or ML demos) to extract its core computational logic, wrap it inside a full-stack SaaS interface (Next.js, FastAPI, Supabase Auth/DB, Stripe/Portone payments), and scaffold cloud infrastructure configurations (Terraform). Use when I say "wrap this github repository into a SaaS", "monetize this open source tool", "convert CLI/script to web SaaS", or "run open-source-to-saas on <repo>".

Analyze successful overseas apps (e.g., Agent job matching platforms) to extract BM/UI/UX elements, adapt them for the local market, and scaffold localized MVPs (mobile frontend, backend APIs, Terraform infra). Use when I mention "clone/localize overseas apps", "research agent job matching platforms", or "run app-cloner-researcher on <app/target>".

Retrieve and rank the top 10 current YouTube trends for a target user segment. Use when the user asks for YouTube trends, trending topics, content ideas, market research, niche discovery, or audience-specific trend analysis for a demographic such as age 40-50, country, language, interests, or creator niche. Always gather fresh evidence before ranking; do not rely on memory.

Download videos, playlists, subtitles, thumbnails, chapters, and metadata with yt-dlp. Use when the user asks to save online video/audio locally, archive a playlist/channel, extract audio, fetch subtitles/transcripts, troubleshoot yt-dlp 403/signature/cookie/plugin issues, configure yt-dlp defaults, or use yt-dlp plugins such as PO token providers, Return YouTube Dislikes, DeArrow, ChromeCookieUnlock, or sample extractor/postprocessor plugins.

Use this skill to generate well-branded interfaces and assets for QuietFabric, a calm technical design system for AI infrastructure (GPU clusters, model serving, distributed systems, observability). Use it for production UI or throwaway prototypes, mocks, decks, and architecture-review presentations. Contains essential design guidelines, colors, type, fonts, assets, UI kit components, and a reusable slide deck.

Summarize YouTube videos from a URL, youtu.be link, Shorts link, embed URL, or raw video ID. Use when I share a YouTube link and ask "요약", "정리", "transcript", "자막", "강의 노트", "영상 내용", "이거 봐줘", or want an Obsidian-ready Korean note with timestamps, key claims, quotes, and follow-up questions. Also use when I ask for "lecture-summarization", "강의 요약", "긴 글로 재구성", or a comprehensive lecture-style article. Fetch the transcript first; do not invent content when captions are unavailable.

Write git commit messages using Conventional Commits. Use whenever I ask to commit, draft a commit message, split commits, amend a commit, prepare a PR commit stack, or say "커밋", "commit", "commit message", "커밋메세지", "커밋 메시지", "conventional commit". Never invent a free-form commit subject when a Conventional Commit type/scope can be chosen.

Prevent image generation models from drawing visible numbers, labels, captions, watermarks, list markers, UI text, or explanatory annotations inside the image. Use when I ask to generate/draw an image and say "번호 빼줘", "넘버링 없이", "텍스트 없이", "라벨 빼고", "글자 넣지 마", or when the prompt contains numbered options that are instructions only and should not appear as visual elements.

Run an autonomous, fixed-budget experiment loop that optimizes one metric: change code → run within a fixed time/step budget → keep or discard by the metric → log → repeat (Andrej Karpathy's autoresearch method). Use when I say "autoresearch", "run experiments overnight", "auto-optimize/auto-tune this", "sweep architecture/hyperparameters autonomously", "keep iterating until <metric> improves", or want an agent to self-improve an ML training script (e.g. nanochat val_bpb) unattended. Establish the metric, budget, edit surface, and a cost/stop ceiling — and confirm them with me — before spending any compute.

Compile a Korean weekly working note from what we did this week. Use when I say "주간노트", "weekly note", "이번주 정리", or ask to summarize the week's work for my Obsidian vault. Output Korean prose, technically precise, no filler.