Recently updated open-source projects

Crux cleans messy transcripts, counts the token reduction, and then guides an agent to remove duplicates, create timestamped chapters, select verbatim quotes, and produce a Markdown pack with an embedded JSON block. It works with YouTube, podcast, or other transcript text; the Python pass is a first cut that still benefits from agent refinement.

AI Monetization Evaluator scores an AI product or side-project idea across demand hardness, niche focus, payment willingness, distribution, and defensibility. It identifies the weakest axis and suggests inexpensive ways to validate that bottleneck before building, keeping attention on demand rather than coding enthusiasm.

visual-diff captures matching pages from a local build and a live or staging URL with Playwright, then compares each pair using odiff. It writes a compact report with a numeric pixel-difference gate, supports desktop, mobile, and full-page captures, and limits image inspection to pages that fail, making visual parity checks cheaper for agents.

A SessionEnd hook records each qualifying Claude Code session to local JSONL drafts or a Notion draft page without blocking your work. Running `/work-log` groups the captured sessions by date and category, summarizes them, and writes a formal log to Notion. It does not require MCP or a separate Anthropic API key, but Notion writes still need an integration token and page access.

react-flow-expert gives Claude Code a version-pinned knowledge base for React Flow, Svelte Flow, and @xyflow. It routes questions to references on architecture, state, nodes and edges, layouts, performance, errors, migrations, and TypeScript types, and can dispatch a focused subagent for implementation or deep Q&A.

linkedin-writer calibrates a user's voice, researches recent industry updates, interviews them for firsthand material, proposes a weekly plan, and writes approved posts with image descriptions. It does not invent numbers, situations, or personal details, and stores the voice profile in VOICE.md for later sessions; a Claude Code account is required.

Collaborative Editor works through structure, grammar, wording, and coherence in that order, raising concrete choices instead of silently rewriting the draft. It treats long sentences, parentheticals, and blunt phrasing as possible voice, then closes with a redline the author can inspect. Outputs include a Word document with native tracked changes plus portable HTML or PDF visual redlines. The workflow is for writers who want control over every accepted change.

nrf-proposal-skill organizes Korean NRF research proposals around program scoring, literature research, research gaps, logic structure, Korean drafting, figures, panel simulation, and HWPX export. It marks unverifiable claims, but the project explicitly treats output as a learning or reference draft that applicants must rewrite and verify rather than submit unchanged.

Give Ultimate Idea Watcher a sentence, rough notes, screenshots, or an existing codebase, and it preserves the original idea while building a structured product brief around it. The skill researches the problem, users, and alternatives, defines the product architecture, then splits work into dependency-aware phases with ready-to-use designer and engineering prompts. It also maintains traceability and gap checks across security, privacy, accessibility, and later plan changes.

Place the scanner between untrusted content, model output, and tool execution, or wrap a line-delimited stdio MCP server with its proxy. It detects prompt injection, credential disclosure, exfiltration, and destructive actions before forwarding a call, redacts secrets, and returns structured reasons for enforcement. The scanner makes no network calls and ships as a Rust CLI with Node and Python wrappers.

Seaworthy performs a static, no-network scan before deployment for common security mistakes in AI-built web apps, including exposed secrets, missing row-level security, unauthorized mutations, risky CORS, debug mode, open admin routes, and writable storage. It returns a ship-or-do-not-deploy decision with file-level findings; it is not a penetration test and primarily targets Next.js and Supabase scenarios.

ecom-static-ads reads one product or collection page, extracts the store’s product imagery and customer reviews, researches customer language and competitor ads, and develops four distinct static DTC concepts. It then generates finished images with the real product photo as a reference and returns a campaign brief, funnel recommendations, an A/B plan, and platform-specific copy. Gemini is required for image generation; research services are optional and fall back to web search.

ecom-klaviyo-email takes one product or collection link and turns it into an editable Klaviyo campaign draft, using brand context, account performance, customer language, competitor research, and a product-referenced hero image. It returns subject-line variants, a preheader, audience prompts, send-time guidance, and local working files, while inserting the store’s reusable header and footer. The audience remains a placeholder and the skill never sends or schedules a campaign; provider keys should be stored outside plaintext dotfiles.

video-bridge retrieves Bilibili subtitles when available and falls back to audio transcription, then turns short caption fragments into timestamped sections, full transcripts, and notes. SQLite FTS5 enables search across videos, with Obsidian-friendly Markdown as an option; subtitle access needs a logged-in session and fallback transcription depends on ffmpeg, yt-dlp, and podcast-bridge.

zx-skills bundles modular skills for Sinclair BASIC, C, Z80 assembly, graphics, AY sound, special display modes, and ZEsarUX control. They support a build-load-run-debug workflow for ZX Spectrum and Timex projects, while actual compilation or emulator control still depends on the relevant toolchains, emulator, and optional MCP server.

lld-flow-explorer reads sequence diagrams, state machines, and flow charts from an LLD, HLD, RFC, or runbook and produces one self-contained HTML explorer with no build step. Readers can choose a flow, play or scrub request hops, inspect wire messages and state machines, and see the active route highlighted on a shared system map. It is intended for documents with several flows through common components; a single linear flow or class diagram is better left static.

sift builds one local full-text index across Claude Code, Codex, and Cursor session files, then returns ranked matches with snippets, timestamps, and tool filters. It can rebuild the index, list recent sessions, and run as a terminal command or an agent skill. Source logs and databases remain read-only, with no network access, telemetry, cloud dependency, or synchronization. It is useful when each coding assistant’s built-in history search is separate or too limited.

HQ Report renders Markdown and frontmatter into a branded PDF with a title page, dark cover, accent headings, severity tables, code blocks, callouts, and page-number footers. A small JSON brand file controls the organization name, badge letter, and color tokens without editing the template, and the same renderer works as a Claude Code skill or standalone Python CLI. WeasyPrint and its native dependencies are required, while the output remains a standard PDF.