ray-os

All LinkedIn tasks — writing posts (text or carousel), checking post performance, browsing the feed, or analyzing competitors. Triggers on any mention of LinkedIn, including "write a LinkedIn post", "make a LinkedIn carousel", "swipeable LinkedIn post", "LinkedIn slides", "check my LinkedIn", "how did my posts do", "LinkedIn competitors", "post on LinkedIn", "browse LinkedIn feed". Also triggers on "check post engagement", "update post metrics", or reviewing social media performance. This is the single master LinkedIn skill — carousel generation, text-post writing, performance checking, and feed browsing all live here.

End-to-end YouTube A/B testing for Ray's channel (@RAmjad) — title generation via subagent, thumbnail image generation via `gemini-3-pro-image-preview` at 1K, results recording, and the underlying scripts/golden references all live here. Use whenever the user wants to brainstorm titles, generate thumbnails (text or images), record A/B test results, analyze title/thumbnail performance, or asks things like "suggest titles", "what titles should I test", "make a thumbnail", "thumbnail variations", "5 thumbnails for this video", "Matt style", "Nate style", "record these results", "what's working", or "update the A/B test results". Also trigger when the user shares a screenshot of YouTube A/B test results and wants analysis or recording. Do NOT use for competitor title research (the youtube-outlier-research routine covers niche/competitor research) or video scripting (use youtube-scriptwriter instead).

Archive Discord channels and forums to local JSON files using Ray's own session — full message history, replies inline, forum threads enumerated and paginated, paced to look like a real browser. Use this skill whenever Ray wants to back up, archive, scrape, dump, or download a Discord server he's a member of, a specific channel, a forum channel's threads, or a server's "questions" / Q&A channels. Triggers on phrases like "archive this Discord", "back up the server", "scrape this channel", "download all messages from #X", "save the cohort questions", "dump the Discord", "make a copy of this server's posts", or any time he points at a Discord URL and wants the content pulled into local files. Also use when he wants to push an archive to a private GitHub repo for safekeeping. Do NOT use this skill for sending messages, moderating, or anything write-oriented — it is strictly read/archive. Honest about what it does: this uses Ray's user token via the same headers his browser sends, which is a Discord ToS grey-zone

Build hand-drawn, Excalidraw-style HTML slide decks that explain a concept through sketch-aesthetic vector diagrams — rough.js strokes on a dark canvas, full-bleed, one diagram per slide with a caption underneath, arrow-key and tap navigation. Use this skill whenever the user wants explainer slides, a concept deck, diagram slides, or a talk/video presentation in the hand-drawn sketch look, or says things like "make slides for X", "build a deck for X", "add a slide in this style", "turn this into a sketch/excalidraw-style deck", "another slide like the workflow deck", or points at an existing rough.js deck and wants more in the same style. This is for self-contained vector-diagram decks — NOT the Claude Code terminal-aesthetic slides (use cc-terminal-slides) and NOT raster image generation (use excalidraw-gen, which calls an image model). Reach for this skill proactively for any "explain this concept on slides" request, even when the user does not say the word "excalidraw".

Reshape an FCP7 XML exported from a silence-cutting app (Recut, AutoEdit, Vika, etc.) so Adobe Premiere Pro imports it cleanly with the user's pristine external audio (e.g. Audio.wav from Audacity) linked to the video instead of the camera/screen-recording's own audio. Use this skill whenever the user points at an .xml or sequence file from Recut, AutoEdit, or any auto-cutter and wants to swap in their own audio track, drop the redundant on-camera audio, fix vertical alignment between video and audio cuts, or fix a Premiere "import failed" error on a Recut-style XML. Triggers on phrases like "fix this Recut XML", "swap the audio in this premiere XML", "align the cuts to my Audio.wav", "remove the loom audio track", "merge V1 and A1 in this XML so they're linked", or any reference to a workflow combining a silence-cutter app's XML with a separately recorded pristine audio file.

Enable a gated, experimental, or staged-rollout Claude Code feature on your own machine by finding its GrowthBook feature flag in the binary and force-enabling it through a Proxyman interception rule (with an optional disk-cache fast-path). Use this skill whenever the user wants to ungate, unlock, force-enable, or "turn on" a Claude Code feature hidden behind a flag — phrasings like "ungate the X feature", "enable ultrareview", "unlock [feature name]", "turn on [feature]", "enable the workflow/agent tool", "this feature won't show up even with the env var", "force-enable a tengu_ flag", "flip the flag for X", "intercept the feature flag", "unlock a beta Claude Code feature", or any case where a Claude Code feature is gated and the user wants it on. Also trigger when the user references a specific tengu_ flag name, or the existing "GrowthBook Flag Enabler" Proxyman rule, or wants to add another flag to it. For the user's own Claude Code install and account.

Search X (Twitter) by topic, keyword, hashtag, or X advanced-search syntax and get back a ranked, deduplicated list of tweets (or accounts) for synthesis. Use this whenever the user wants to discover what's being said about a topic on X without already having tweet URLs in hand. Trigger on phrases like "search X for X", "search Twitter for X", "what is being said about X on X", "find tweets about X", "pull recent X discourse on Y", "any X chatter about Y", "scan X for Z", "what are people tweeting about X", "is anyone on X talking about Y", or any request to discover, sample, or harvest X content on a topic the user does not already have URLs for. Also use it proactively inside multi-agent workflows where one agent needs fresh X signal on a topic before generating ideas, drafting copy, doing competitor research, or sizing a hot take. Distinguished from `x-thread-miner`: if the user hands you a tweet URL, use x-thread-miner to mine its replies and quote tweets; if the user hands you a topic, a keyword, a hasht

Fetch the replies and/or quote tweets for one or more individual X/Twitter posts, then mine them for OUTLIER ideas — non-obvious takes, contrarian theses, sharp objections, and novel framings — and render them into a navigable HTML artifact with pin/group/export. Use this whenever the user points at one or more tweets (URLs or ids) and wants to know what people are saying back, gather the reactions, harvest the replies and quote tweets, see how a feature or idea is being received, or turn viral threads into research. Trigger on phrases like "mine this tweet", "get the replies and quotes for this", "what are people saying under these posts", "pull the quote tweets", "gather reactions", "what's the discourse on this", "find outliers", "what's the non-obvious take", or any time the user shares tweet links and wants the conversation around them distilled rather than just the tweets themselves. Always quiz the user first on whether they want quote tweets, replies, or both. Use this even when the user does not say

Write the short X/Twitter post that sits ABOVE an embedded YouTube video to make people watch it, in the proven "video-tweet" style reverse-engineered from Matt Pocock's highest-performing video posts. Also logs each published post and checks its performance a few days later against the benchmark. Use this skill whenever Ray has just published (or is about to publish) a video and wants the tweet/caption to go with it: "write a promo tweet for this video", "tweet to go with the video", "X caption for my video", "video tweet", "short post for the video below", "promo post for [video]". Also use it to record a published promo post ("log this promo tweet") and to grade how one did ("how did the promo tweet do", "check the promo post performance", "did the video tweet land"). This is for SHORT scroll-stopping posts attached to a video, NOT long-form articles (use x-article), NOT reply/quote research (use x-thread-miner), and NOT LinkedIn (use linkedin).

Generate excalidraw-style explanation images using the Codex CLI (codex exec --json) instead of a paid API. Use when the user wants visual explanations, diagrams, or illustrations in the hand-drawn excalidraw aesthetic and asks for the codex path. Codex's bundled image_gen tool runs through the logged-in account, so no GEMINI_API_KEY is required and it's often faster than the Gemini route. Triggers on "generate an excalidraw image", "excalidraw-codex", "use codex for excalidraw", or any request for excalidraw-style diagrams where no engine is named. This is the RECOMMENDED DEFAULT path for excalidraw image generation: it needs no API key and runs through the logged-in Codex account. Sibling skill [[excalidraw-gemini]] does the same job through the paid Gemini API. Default to codex whenever the user does not explicitly ask for the Gemini route.

Curiosity engine for JAPANESE-immersion video discovery — "what Japanese video should I watch next?" Harvests candidate Japanese videos from YouTube's unauthenticated graph (related-video rails around what you loved, JP-native search, fresh uploads from your channels), then YOU (this session — no cloud LLM) judge, rank, and diversify them against the taste on record. Bare `/watch-next-jp` does an open curiosity pass (breadth across genres, novelty-weighted); `/watch-next-jp about <topic>` narrows the region but keeps the variety. Reads taste.json (your watch history + channels) as seeds; writes nothing to it. Use for "/watch-next-jp", "/watch-next-jp about X", "what Japanese video should I watch", "recommend me something to immerse in", "find me new Japanese channels", "what should I watch for immersion tonight". ONLY for Japanese-language immersion viewing — do NOT use for general/English video recommendations, for Ray's own YouTube channel (that is youtube-ab-tester / youtube-outlier-research), or for any n

Turn ONE video (YouTube URL/id or a pasted transcript) into an Agentic Coding School (ACS) content-gap report: extract its full wisdom, promote the 1-3 load-bearing "spine" ideas, deep-dive each in prose, gap-check each against the live ACS catalog, and pitch the buildable new ACS videos. The analysis engine behind the acs/gap-scout routine, but also runs standalone. Use this skill WHENEVER the user pastes or links a coding/AI video (podcast, talk, tutorial, feature walkthrough) and wants the ACS angle: "what ACS video could come from this", "run the gap scout on this", "gap-check this against the catalog", "find the spine idea", "is this already covered in the school", "what should I film from this", "mine this video for ACS", or just drops a transcript/link and asks what's video-worthy. Also the per-video worker the gap-scout routine calls for each swept video. Do NOT use for a bulk channel sweep (that is the gap-scout routine / youtube-outlier-research routine) or for plain wisdom extraction with no ACS ga

Mine a YouTube creator's entire back catalogue (or a date-bounded slice) for reusable ideas, content angles, and insights, tracing each insight to the original source it came from (blog, paper, tweet, changelog) and surfacing what the audience keeps asking for in the comments. Built on VidTempla (channel video listing + comment threads) and Supadata (transcripts + descriptions). Use this skill WHENEVER the user wants to go through a channel's videos in bulk to extract ideas, not just look up one video: "mine Theo's videos", "go through every video on this channel", "what video ideas can I pull from <creator>", "analyze <channel> for content angles", "what is <creator>'s audience asking for", "find the best insights across <creator>'s last year", "scan <competitor>'s channel for gaps", "turn this creator's catalogue into a list of ideas". Trigger even when the user names a creator/handle/URL and says "find me ideas" without saying "mine" or "skill". Produces a ranked, source-traceable markdown report plus an i

Turn Ray's messy notes, memos, transcripts, or half-formed ideas into a filmable YouTube video for his main channel (@RAmjad), as an image-first bullet deck he riffs over (NOT a prose script). The workflow is always three phases in order: interview first to develop and stress-test the idea, then write the deck, then generate the images. Use this skill whenever Ray wants to make a video, write a YouTube script, turn a topic or memo or artefact into a video, plan video beats, or says "/youtube-writer", "write this up as a video", "let's make a video about X", "turn this into a YouTube video", or drops a topic clearly meant for the channel. Prefer this over free-handing a script: the interview is what makes the video good, and skipping it produces thin, generic decks.

Write personal newsletters in a conversational, essay-style format inspired by Ali Abdaal's style. Use when the user wants to write a newsletter issue, draft an email to their list, write a LifeNotes-style essay, or asks for help with newsletter writing. Triggers on requests like "write a newsletter", "draft a newsletter about...", "help me write my weekly email", "newsletter style", or when the user provides a topic and wants it turned into a newsletter format.

Generate excalidraw-style explanation images using the Gemini API (gemini-3-pro-image-preview). Use when the user wants to create visual explanations, diagrams, or illustrations in the hand-drawn excalidraw aesthetic via the Gemini path. Supports parallel generation of multiple variations with configurable timeout. Triggers ONLY when the user explicitly asks for the Gemini route ("generate an excalidraw image with gemini", "excalidraw-gemini", "use gemini for excalidraw"). Sibling skill [[excalidraw-codex]] does the same job through the `codex exec` headless CLI with no API key, and is the RECOMMENDED DEFAULT. Prefer codex unless the user explicitly names Gemini.

Transcribe audio or video into a clean .txt + timestamped .srt + YouTube chapter timestamps. Re-encodes any WAV/MP4 to a small MP3 before upload, mines on-screen text with OCR to seed AssemblyAI key terms, runs AssemblyAI Universal-3 Pro speech-to-text, then generates punchy chapter markers from the SRT. Use whenever the user wants to transcribe a video or audio file, generate subtitles/captions or an SRT, create a YouTube transcript or chapters, or convert a recording to text. Also trigger when the user has a WAV/MP4/MOV and wants text out of it, even without saying 'transcribe'.

Turn a raw idea, source tweet, video transcript, feature announcement, or rough outline into a polished video script for Ray's Agentic Coding School classes. Use this skill whenever the user says "write a class script", "turn this into a class video", "script this for the [business/claude-code/prompt-engineering/skills/context-engineering/techniques/claude-chat/claude-cowork/correction] class", "draft an agentic coding school video", or whenever the user is staging content in `/Users/ray/Desktop/ray-os/projects/agentic-coding-school/to-film/` and wants it written up. Default to this skill for any new file destined for the `to-film/` tree, even if the user does not name it explicitly. Produces an essay-style markdown script with the correct frontmatter, horizontal rules between sections, inline image embed placeholders, and a Demo section, written in Ray's direct second-person voice.

Reverse-engineer any desktop application OR Claude Code's own binary to understand how it works behind the scenes — the APIs it calls, the internal flow, the models/providers it uses, its feature flags, config constants, embedded prompts, IPC protocols. For general apps it works on Electron apps (Wispr Flow, Slack, Discord, VS Code, Cursor, Notion, Linear…), native macOS .app bundles, single-executable bundles (Bun/Node/Deno/Go SEA), and raw Mach-O / ELF / PE binaries. It ALSO reverse-engineers Claude Code's own binary — it knows Claude Code's mangled variable conventions, feature-flag patterns, and embedded-prompt layout. Use this skill whenever the user points at a desktop app and wants to know "how does X work under the hood", "what API does it call", "how does it do Y", "what's the flow for Z", "reverse engineer this app", "figure out how this feature works", "what endpoints does it hit", "does it run locally or in the cloud", "what model does it use", or names a specific app path (.app, .exe, a binary).

Extract structured wisdom, insights, and actionable takeaways from any content — podcasts, articles, videos, transcripts, books, or conversations. Use this skill whenever the user asks to "extract wisdom", "summarize insights", "pull out key ideas", "what are the takeaways", "extract the best parts", "distill this content", "what did I learn from this", or provides a transcript/article and asks for a structured breakdown. Also trigger when the user pastes or references long-form content and wants the signal extracted from the noise — even if they don't use the word "wisdom" specifically. Inspired by Daniel Miessler's Fabric pattern. Has a BATCH MODE for running this extraction across a whole YouTube channel in bulk: lists videos with yt-dlp, filters by duration, downloads transcripts, then fans out one extract-wisdom subagent per video in batches of 5, writing one wisdom file per video to disk. Trigger batch mode whenever the user wants wisdom/insights/takeaways from a YouTube channel in bulk: "extract wisdom

Remove signs of AI-generated writing from text and revise AI-like, generic, over-polished, promotional, or detector-looking prose into clearer human writing. Use when asked to humanize, de-AI, make prose sound natural, reduce chatbot style, remove AI tells, improve voice, rewrite drafts, edit Wikipedia-style text, polish essays/posts/emails, or diagnose AI-ish writing patterns — while preserving factual accuracy, citations, author intent, and appropriate disclosure. Based on Wikipedia's comprehensive "Signs of AI writing" guide. Detects and fixes patterns including: inflated symbolism, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary words, negative parallelisms, and excessive conjunctive phrases. Does NOT help with deception, plagiarism, fabricated sourcing, or evasion of required authorship/disclosure rules.

Audit and clean up an Anki deck the user is struggling with — find the cards that aren't sticking, surface the HIGH-LEVEL PATTERNS behind them (not just a flat list), triage the problems into prioritized buckets, fix the bucket the user picks, then loop back and re-offer the rest. Use this whenever the user wants to tidy up, audit, clean up, fix, prune, reprioritize, or "do something about" an Anki deck or their spaced-repetition reviews — phrasings like "clean up my deck", "what am I struggling with in Anki", "my leeches are out of control", "audit my sentence mining deck", "fix the cards I keep forgetting", "tidy up my reviews", "reset my hard cards", "what's wrong with this deck", or "help me get my Anki under control". Also trigger when the user points at AnkiConnect / a deck name and asks for analysis or maintenance, even if they don't say the word "cleanup". For finding/replacing a single bad example sentence or mining new cards, defer to the sentence-mining skill; this skill is the deck-wide audit-and-

Write a long-form X/Twitter article from a YouTube video, transcript, topic, or agentic coding school content. Fetches the transcript, distills core ideas, and writes a platform-native article optimized for X's long-form format. Can combine multiple sources — school videos, external X posts, research — into a single cohesive article. Use this skill whenever the user wants to write an X article, turn a video into an X post, or says things like "write an X article", "X article from this video", "long-form X post", "write this up for X/Twitter". Also triggers on "article for X" or "post this on X" when long-form content is implied. Triggers on "combine insights" or "daily article from school" for scheduled runs.

Mine 60 to 150 fresh ideas on any topic, concept, project, video, class, or research question by fanning out ~20 parallel agents across web search (Exa), X discourse (via the x-search skill), and 8 pure-reasoning angles (economics, analogies, failure modes, cross-domain, dark side, first principles, anti-patterns, provocations). Synthesizes the raw pool, dedups, ranks by novelty, and renders to a self-contained dark-theme HTML idea bank with selectable cards, a floating "Copy selected (N)" button, filter chips, search, and keyboard shortcuts. Use this skill whenever the user wants to brainstorm widely before narrowing, mine angles on a topic, explore the design space, generate ideas, build an idea bank, research a concept from many sides, stage ideas to curate later, prep for a video series, prep for a class, or kick off a new project with a rich raw material pool. Trigger on phrases like "mine ideas for X", "brainstorm Y", "idea bank for", "generate ideas about", "what could I do/say/build/write about X", "g

Bundle a set of section images into a single .excalidraw file laid out as a grid — one column per section in document order, the section's variations stacked vertically — and save it to ~/Downloads, ready to drag into excalidraw.com or the Obsidian Excalidraw plugin.

Build Japanese sentence-mining cards for Ray's Anki deck in two modes. (1) Video mode — paste any Instagram reel, YouTube video/Short, TikTok, Twitter video, or local file → yt-dlp + AssemblyAI + mecab + AnkiMorphs i+1 diff produces draft cards. (2) Bank mode — give a list of target words → search across Ray's locally-indexed subs2srs .apkg banks (Tokyo Ghoul, Harry Potter, LOGH, etc.) for natural example sentences, reusing the bank's original audio + screenshot when available. Both modes stage drafts for Ray's approval before pushing via AnkiConnect. Use proactively whenever input is (a) a Japanese-language video URL or (b) a list of Japanese words — don't ask if Ray wants cards, start drafting. Trigger phrases include "mine this video", "make sentence cards from <url>", "turn this reel into cards", "mine these words", "find sentences for [w1, w2, …]", "i keep forgetting <word>", "pull cards from my <show> bank", "leech these", "search the banks for X", `/sentence-mining`, or any video URL paired with a ment

Orchestrator skill — takes a YouTube video and cascades it across platforms by calling other skills. Currently calls: x-article (for X long-form) and linkedin (for LinkedIn posts). Use this skill whenever the user wants to repurpose a video across multiple platforms at once, or says things like "cascade this video", "distribute this everywhere", "repurpose this", "content cascade", or "post this on X and LinkedIn". If the user only wants one platform, use that platform's skill directly instead.

Clean up unreferenced images in Obsidian vault folders. Deletes unused images, renames excalidraw files with descriptive names based on image content (using vision), and moves everything to vault root images/ folder. Use after picking favorites from excalidraw-gen batches. Triggers on "clean up images", "delete unused images", "tidy up images", "image cleanup", or "/image-cleanup".