Shape agent capabilities into your workflow.

multi-model-debate-skill runs a live panel in which each model argues a distinct lens instead of repeating the same safe answer. Round one is blinded, genuine objections are required before agreement, and a devil’s-advocate control can challenge premature consensus. Opus moderates the running session while Grok, Codex, DeepSeek, or other configured backends supply additional voices; the lineup can be chosen per debate.

fiz-invoicing-skill uses the FIZ API to find or create customers and products or services, apply the appropriate VAT rate, and build a draft invoice with discounts or payment details. You then issue the document—the legally binding step reported to the Portuguese tax authority—and download its PDF. Claude Code, a FIZ account, and an API key are required.

timplay moves product planning through four dependent stages: a question-driven PRD, an editable feature tree, a user-flow diagram, and a clickable wireframe prototype. The HTML tools support drag-and-drop, node and connection editing, responsive desktop/mobile views, and JSON round-trips for browser edits. Generated product content follows the input language, but the surrounding HTML controls are currently fixed in Korean.

claude-telegram-bridge gives a long or unattended Claude run a Telegram side channel. notify sends non-blocking progress, while ask or listen waits for a reply and exposes clear exit codes for answered, timed out, unconfigured, or stopped states. Replies are accepted only from allow-listed chat IDs, so another person cannot drive the agent through the bot.

PaperRadar sweeps enabled preprint, journal, and open-repository sources for the past week, then ranks papers against your topics and arXiv categories instead of raw popularity. It writes a ranked weekly note and linked per-paper scaffolds with abstracts, links, and citation IDs, with optional dated Zotero filing. The skill runs without a database, embedding service, or web app, but needs its Python dependencies.

ai-equity-analyst reads only the investor presentations and earnings-call transcripts you provide for an Indian listed company. It applies a 33-question checklist, distinguishes hard signals from soft guidance, and produces bull and bear cases with passage-level citations plus a promise-versus-reality ledger. It is for research and education, not SEBI-registered investment advice, and it does not fill gaps from memory.

teams-video-dl-skill extracts a Teams recording and its speaker-labeled transcript from a signed-in Chromium browser session, then uses ffmpeg to mux the video and audio tracks. Scripts run through uv, so no manual Python dependency setup is needed. Processing stays local; temporary HAR and curl files can contain session cookies and tokens and are removed after a successful run, making the local machine boundary important.

ralph-loop-setup wraps the Ralph loop for Claude Code or Codex: write a PRD, split it into atomic stories, and let a fresh agent implement one unfinished story at a time with checks and commits. Progress persists in prd.json, progress.txt, and Git history. A branch is not done until both the L1 PRD-focused review and an independent L2 model review are green, or every finding is explicitly closed.

youtube-to-text accepts a YouTube URL or local file, extracts audio, and transcribes it with local faster-whisper or ElevenLabs Scribe v2. It outputs paragraph-segmented text, optional SRT subtitles, and an optional summary; Claude can clean up English brand and technical terms in Hebrew transcripts. Apple Silicon macOS is the primary target, with local processing by default, while the cloud engine requires an ElevenLabs key.

repo-proof runs a repository’s quickstart inside a sandbox, fixes documentation when commands are wrong, and verifies the result again before capturing a real demo. It refuses to hide an actual failure behind polished copy, making it useful for maintainers who want to demonstrate that a Claude Code skill or open-source project really works. The process executes project commands in an isolated environment.

decard finds decorative card containers in shadcn, MUI, Bootstrap, Vuetify, Ant, Chakra, or custom UI and replaces them with sections, headings, and dividers. It preserves structural frames around tables, lists, code blocks, and inputs, and can collapse repeated card grids into divided lists. Use it to reduce visual wrapping without removing meaningful boundaries.

geo-content-gap compares where an AI engine retrieves your pages with where it actually cites them for each tracked prompt. It lists your retrieved or cited URLs, the third-party sources being cited, and routes each prompt to no action, existing-page optimization, or new content. The output includes a prioritized backlog and optional briefs in Markdown, Google Sheets-ready TSV, and Excel fallback formats, using whatever visibility source supplies the required per-prompt data.

carousel-maker turns an article, X thread, or reel into an 8–10-slide Instagram carousel. It gathers source text and visuals, designs a topic-fitting palette and story, builds a 1080×1080 HTML deck, and renders static PNG or motion MP4 files in post order. It asks for brand choices on first use and can iterate on feedback; reels require yt-dlp, while rendering uses ffmpeg and Playwright with Chrome.

Butler treats the main agent as the execution owner: it understands the goal, makes key judgments, integrates results, and performs final acceptance. It may delegate bounded, verifiable subtasks to a lower-cost model only when the total effort is worthwhile, and it can choose to do the work directly instead. Delegated output must be checked by the main agent, with explicit terminal states for completion, decisions, or a new work window.

gallup-hr-skills packages two complementary views of Gallup’s evidence-based HR framework: one for culture and HR teams designing or interpreting engagement surveys, and one for managers working with their own teams. Both share Q12, CliftonStrengths, the 70% manager lever, and the evidence-oriented framework; the installer lets you choose the culture/HR view, the manager view, or both. It is a methodology advisor rather than a roleplay persona.

design-styles is a library of 50 named visual directions, from Neoclassical and Japandi to Synthwave and Glassmorphism. Each entry pairs a reference image with core elements, mood, layout and motion notes, starter CSS tokens, and an adaptable image prompt. Use it to compare a few directions, build a moodboard, combine styles, or brief an image, video, carousel, or frontend concept.

SSW.YouTubeWatchStats uses a dedicated persistent browser profile to read roughly 30 days of YouTube history and write a developer-focused Markdown report. It separates technical from non-technical viewing, groups topics and formats, and estimates watch time and leading channels. History stays on the machine, but the skill requires an interactive signed-in browser window, so it will not run in a headless environment without a display.

berjiwa is a Claude skill for Malaysian Malay marketing and advertising copy. It combines emotion-first storytelling with a humanizer that removes common AI patterns and detects Indonesian intrusions, covering short-video scripts, sales pages, ads, social captions, emails, headlines, and hooks. The workflow is intended for teams that want persuasive copy with a more local, human voice.