Recently updated open-source projects

imessage-spam-cleanup scans the Messages chat database read-only, classifies conversations as spam, normal, or unknown, and creates a human-readable report before anything is removed. After explicit approval, it uses Messages.app UI automation, protects whitelisted contacts, and relies on the app's Recently Deleted area for 30-day recovery. The skill requires macOS 14 or later, a signed-in Messages.app, Full Disk Access, Accessibility permission, and Python 3.9+.

This collection packages three skills: a triaged six-step work process with an approval gate, a structured evaluation for external repositories, and a benchmark process for choosing production AI models. Its accompanying CLAUDE.md adds rules for evidence labels, security scope, parallel-agent isolation, and commit discipline, aimed at reducing scope drift and unapproved execution.

VirtualSMS exposes physical-SIM phone numbers for receiving verification SMS and OTP codes across more than 145 countries and 2,500 services. The Agent Skill can be used from Claude, Cursor, Codex, Windsurf, OpenClaw, and other MCP-aware clients through an MCP server, REST API, or x402 pay-per-call endpoints. Pricing is returned at request time rather than hardcoded; MCP use requires the service configuration and an API key, while x402 uses USDC on Base.

research-agent is a set of three system prompts rather than a standalone program. A lead agent plans the question and delegates up to 20 parallel workers, research subagents gather facts within a tool budget, and the lead synthesizes their results into Markdown before a citations agent adds attribution without changing the prose. It is intended for hosts that can run subagents and research tools, and the prompts enforce source-quality and role boundaries.

Timeln Second Brain lets Claude Code or Cursor answer questions from the documents and relationships stored in your Timeln library. Through its MCP tools, an agent can inspect recent material, query connected topics, retrieve documents, and ingest text or public URLs; the hosted setup needs a Timeln account and bearer token, with a local Python MCP available as an alternative.

cm-committee routes a methodology question to a structured debate among seven independently deployable scholarly perspectives, including systems dynamics, urban systems, spatial statistics, Bayesian workflow, and leverage-point thinking. Each lens is grounded in primary sources, states where it should defer, and can be used alone or orchestrated with others. The result is a disciplined thinking aid, not an authentic statement from any scholar and not a substitute for factual research.

This Claude Code skill interviews a patient or caregiver, then creates a long-running research repository with an operating contract, evidence standards, literature methods, treatment records, mechanisms, clinician notes, and open questions. A required frontier scan must be completed before candidate-treatment work is allowed. It is for people already working with a specialist and is not a symptom tracker, diagnostic tool, prescription writer, or substitute for clinical care.

Generate a voicemail greeting from flags for message type, industry, name, phone number, and business hours, or supply your own script. The skill keeps drafts to about 30 seconds, then links to Solvea to choose from six AI voices, preview the result, and download an MP3. Free Solvea accounts include up to 13 generations.

Game Architect helps solo developers and small teams turn a concept into structured design decisions across mechanics, balancing, narrative, platform and engine choice, audio, Steam marketing, and post-launch operations. It uses genre presets and can scaffold GDD and Claude Code project documentation. Its code material is limited to simulation tools, illustrative pseudocode, and adaptable templates; implementation decisions remain with the developer.

prompt-refiner audits prompts you have already written in Claude Code, Cursor, or ChatGPT against a documented catalog of clarity, output-hygiene, and prompt-injection rules. A zero-dependency Node detector handles reproducible checks and can run in CI, while semantic cases use the active model. It reports findings with locations and evidence, but does not rewrite the prompt, impose a framework, or interview the user.

Use ultra-plan when a feature spans several modules or requires multiple AI sessions. It probes project rules, dispatches parallel read-only research, builds a module dependency graph, writes a global plan plus one document per module, and then asks how to execute. Separate Claude Code and Codex versions adapt the prompts and dispatch tools; routine single-file work is intentionally out of scope.

Running `/bro` stores repository-specific context that chat compaction can erase, including rules, architecture decisions, vocabulary, roadmap changes, and active debugging hypotheses. It separates persistent principles from thread state and daily logs, tags parallel chats with the title plus a UUID suffix, and can migrate older log layouts into the visible project `bro/` directory.

Coordinate discovery, design, architecture, implementation, testing, release, and operations with schema-validated artifacts and script-enforced transitions. Implementation and test gates run declared verification commands rather than trusting filled-in status files, while upstream revisions mark downstream work stale. Key risk gates require confirmation, and the implementation phase depends on the separate harness-plan skill.

Use this skill when a Midnight wallet stalls, a deployment hangs, the wrong address is selected, an RPC endpoint is blocked, or a Groth16 verifier may not match its local key. It maps symptoms to focused runbooks and includes scripts for address derivation, reachability checks, hash parity, and post-deploy verification. It is intended for developers operating Midnight Network backends through Claude Code, Codex, Cursor, or another markdown-aware agent.

Visa Application Skill guides an application from visa-type intake through current official-policy research, difficulty assessment, tiered checklists, fact-based document drafting, file compliance checks, and a submission walkthrough. It can support a guided browser workflow or semi-automated form filling with Chrome MCP, but it never clicks the final Submit or Pay action. Missing facts stay visible as questions, policy claims are sourced at runtime, and applicants must recheck official guidance before filing; the tool is not legal advice.

dark-tech-style gives Claude a fixed visual language for architecture pages, dashboards, roadmaps, technical landing pages, and one-pagers. It includes four dark background levels, seven accent colors with dim variants, typography and spacing tokens, a glassy navigation treatment, and 14 component classes, plus starter HTML and demos. The skill emphasizes one accent per section, background-only dim colors, and token-based styling instead of hardcoded hex values.

Import a PDF by rendering every page as a JPEG and inserting the images as full-width blocks in a new Notion page, creating a presentation-like scrolling view. The command can set the title, reading status, and tags and can add generated cover and icon artwork. A Notion token and database ID are required; an Unsplash key is optional for cover images.

Bridge findings from Cowork, Claude.ai, Claude Desktop, or Claude for Work into Claude Code through a shared markdown inbox. The /cowork skill classifies scope, asks for one confirmation, runs in-scope items sequentially, archives completed findings, and produces an aggregate report. Data-destructive work always pauses for explicit confirmation, and upstream inbox population remains your choice.