Recently updated open-source projects

brand-coherence-audit compares a brand or person across up to 15 public web sources, extracting claims about positioning, offers, expertise, and proof. It produces weighted scores, a cross-source comparison matrix, quoted inconsistencies, and a prioritized update plan aimed at clearer AI-search understanding. The skill orchestrates external collection tools, so API access may be needed for social and search sources.

OCULUS Design System packages the company's visual and writing guidelines as a Claude skill for documents, presentations, print, web, and app interfaces. It includes the O-Blue palette, Eurostile and Segoe UI typography, logo spacing rules, text notation, CSS/JSON tokens, and assets. Brand assets remain the property of OCULUS and require authorized use.

Council sends the same review prompt to several configured AI coding agents in parallel, then asks the host agent to synthesize a deduplicated report for staged Git changes. It can mix Claude Code, Codex, OpenCode, and Cursor backends, but requires at least one supported and authenticated CLI.

This Claude Code skill turns package targets, CVE advisories, Jira tickets, or Dependabot alerts into an upgrade workflow for Python, JavaScript/TypeScript, and Go. It analyzes direct and transitive dependencies, reviews breaking changes, proposes confirmed code edits, runs layered tests, and can finish with commits, pull requests, and Jira updates.

security-audit guides Claude Code through a defensive review of a codebase, an authorized black-box target, or both. It selects the appropriate checklist for authentication, injection, SSRF, file handling, infrastructure exposure, and rate-limit posture, then emphasizes verification and coordinated disclosure. The runbook assumes explicit permission for external probing and is not a substitute for a funded penetration test or a guide for unauthorized targeting.

vibe-prompt interviews you about an app idea, then produces a concise spec, an initial plan-mode build prompt, an ordered sequence of independently verifiable feature prompts, a workflow checklist, and a short rationale. It encourages one vertical slice at a time with real-app verification. The plan-mode workflow is tailored to Claude Code, even though the prompts can travel elsewhere.

Turn Follow Builders' text digest into a newspaper-style HTML daily with a masthead, headline, quick scan, quote, builder cards, and podcast feature. It adds presentation and layout rather than sourcing the data, so the original Follow Builders skill is required as the content source.

env-paper-writing collects six Claude skills for environmental science and engineering papers: section drafting, abstracts, figure captions, cover letters, reviewer responses, and data-availability statements. It preserves LaTeX commands, uses domain terminology and units, and avoids inventing missing data; authors remain responsible for accuracy and disclosure.

watch-video uses yt-dlp to fetch videos or accept local files, takes captions when available, falls back to local whisper.cpp transcription, and extracts adaptive video frames with ffmpeg. The agent receives metadata, a transcript, and a frame index so it can inspect only the visuals needed for an answer, with fast, balanced, and accurate modes controlling cost. No API key or third-party transcription service is required, but local tools and an optional Whisper model are needed; Claude Desktop is not supported.

jira-planner guides a free-form problem through research, estimation, structured ticket drafting, user confirmation, creation, and field verification. It reads and writes Jira Cloud through an Atlassian MCP server, with the optional acli tool used for faster read operations. The skill is tenant-agnostic, but it remains inactive until its local site and field configuration has been set up.

Deck Pipeline takes a Chinese presentation through design analysis, page-by-page English rewriting, layout checks, and a documented handoff. It produces edited PPTX files and a bilingual comparison workbook, while a polish-only mode can clean up single-language decks without translating them. Slide checkpoints let users correct terminology or direction before changes spread across the deck.

Give it a PRD or product specification, and it chooses between a compact promise-based plan and a broader feature-grid approach before drafting P0, P1, and P2 scenarios. Each scenario pairs user-interface checks with backend assertions and accounts for multi-role permissions, making the skill useful for planning regression coverage in SaaS and other complex products.

offboarding-auditor examines a departing engineer's code and documentation for knowledge silos, account bindings, unfinished work, dependencies, and test coverage. It scores how independently a new owner could take over, then produces a severity-ranked report and assignable fixes to complete before departure. The focus is inheritability, not code style.

This skill uses LaMa-based inpainting to remove corner watermarks, logos, date stamps, microphones, and other tracked objects from MP4 videos while leaving audio untouched. It can process folders and skip completed files, but is tested on macOS, needs ffmpeg plus Python and a model download, and cannot handle large-area or audio watermarks.

Magic Memory coaches a learner through calibration, intuition-building, prediction, retrieval, and flashcard creation instead of stopping at a polished explanation. Concepts that stick become cards scheduled by a standalone SM-2 Python engine. It suits abstract mechanisms, programming ideas, architecture tradeoffs, and confusing pairs of terms.

Use slidoc to extract slide frames from lecture or training videos, transcribe the audio locally with whisper.cpp, align narration to each frame, and produce searchable Markdown notes. You can run its Python CLI stage by stage or let the bundled Claude Code skill drive the workflow. It requires ffmpeg, whisper.cpp, and a local Whisper model, and transcription of long batches can still take several hours.

Provide a website URL and Agentizer crawls its brand and product content, then generates a Next.js agentfront with chat, voice, interactive browsing, and conversational conversion flows. It can run through Claude Code, Codex, or a standalone CLI and deploy through Vercel and Cloudflare. The generated experience requires Gemini and OpenAI API credentials.

SUPER evaluates whether LLM agents can configure environments, execute code, and answer questions drawn from real machine-learning and NLP research repositories. Its dataset is divided into Expert, Masked, and AutoGen task sets, with code for running agents and scoring results. Local mode executes repository code directly on the machine, while Docker and Modal backends offer safer isolation and concurrent benchmark runs.