Recently updated open-source projects

wh-skill records new commits through a global post-commit hook and turns them into Markdown work-hour reports. You can filter by day, week, month, date range, or project, switch to chat-friendly text, group totals by project, and add meetings or reviews manually. Tracking begins only after installation, so earlier commit history is not backfilled.

Verify by Execution requires reproducing the problem before changing code and rerunning the exact scenario afterward. Every conclusion includes the command, relevant output, and a pass, fail, hang, or skip result. Its rules also cover full test suites, clean-state checks, both sides of security gates, and unset, empty, valid, or invalid environment variables.

Adversarial Audit executes the repository's claimed test and smoke commands instead of trusting source inspection or a zero exit code. It checks skipped tests, weak assertions, swallowed failures, hanging processes, and both the blocking and allowing sides of security gates. The skill is aimed at finding false-green CI and verification gaps that reading alone misses.

Describe a scenario in plain language and the skill follows it from the entry point through every real call site. It tracks how values, formats, and units change, then flags stale credentials, missing consumers, partial updates, unit mismatches, or swallowed errors. This is useful when each file looks correct alone but the full interaction fails.

Seven-Lens Review makes Claude examine a target through contracts, state transitions, boundaries, security, cross-component consistency, dead interfaces, and test gaps in one pass. The checklist is language-agnostic and works best on features spanning several files, where individually reasonable components can still disagree. It is a single skill file with no added dependencies.

Describe the systems you already run and the skill maps where AI could add value, compares integration patterns, and generates starter code for MCP servers, API connectors, RAG pipelines, middleware, or event-driven agents. It also covers deployment, monitoring, security controls, cost estimates, and phased rollout planning. The focus is early architecture and implementation scaffolding for enterprise teams.

Project Navigator searches your Claude conversation history for decisions, prior deliverables, evolving ideas, and current project status. Its pattern mode can extract the structure or method used in earlier work and present it as a reusable template before applying it again. It uses Claude's built-in search rather than external storage, so results depend on what those tools can surface and cannot restore old files directly.

Roast challenges an idea, feature brief, or rough concept instead of defaulting to validation. It examines whether the problem is real, whether the value justifies the complexity, what the product would lose, and what conditions must be met to proceed. The result is a direct build, do-not-build, or major-rethink verdict for PMs, designers, and builders.

In Create mode, DesignLint makes the agent understand the user, choose a creative tension, and commit to nine decisions across aesthetics, typography, color, space, motion, interaction, voice, and departure before implementation. In Audit mode, it documents the current UI, scans for repetitive AI patterns, scores the design, and proposes the three highest-impact improvements. It supports Claude Code, Cursor, Codex, and other agents.

Bazaar treats GitHub repositories tagged `claude-skill` as discoverable skills and exposes local commands for searching, viewing, installing, publishing, and following creators. It uses an authenticated GitHub CLI underneath, so there is no separate backend or marketplace account to maintain.

Agent Memory organizes Markdown notes into a searchable vault with YAML metadata, relevance decay, and core-to-archive tiers. Agents can load only a description, a cached summary, or the full card, search different tiers by task depth, compress old session files, and inspect context-budget usage. Core memories remain manually controlled, and the basic engine has no required external dependency.

A PostToolUse hook records tool calls that were not covered by existing allow rules and therefore required manual confirmation. The analyzer filters those entries by project, groups repeated command patterns, and presents suggested permission rules with examples. Nothing is added automatically: the user chooses which proposals should be written to Claude’s settings.

betting-app-skill is a Claude skill for building sports-betting and prediction-market applications with Next.js and Supabase. It documents PostgreSQL schemas, RLS policies, an atomic bet-placement function, pari-mutuel odds, settlement, leaderboards, and fixes for stale data or join types. The repository is a development guide extracted from a reference application, not a hosted betting product or standalone wagering service.

The skill gives Claude Code reference material, reusable TaxEngine patterns, and calculation checks for SST, GST, and VAT implementations. Its strongest documented coverage is Malaysia and Singapore, while the Thailand material remains a skeleton. Income, payroll, and corporate taxes are outside scope, and production rates still require official or professional verification.

The skill reads an application’s code, blueprint, or description, then identifies missing features, useful AI integrations, n8n automation opportunities, and library changes. It organizes the findings by priority and can generate detailed specifications for selected improvements, with implementation remaining an optional next step.

This skill first captures a business idea as a brief, then researches competitors, and finally creates a development blueprint with business and technical PRDs, database schemas, API contracts, module specifications, testing guidance, a multi-agent plan, and an implementation roadmap. The output is a structured set of documents and project scaffolding for later coding, not a finished application. Web search, Node.js with docx, and Python with openpyxl are required.

The Dock0 skill guides developers through connecting an existing remote MCP server or deploying a stdio server from GitHub onto managed infrastructure. It covers endpoint setup, API keys, rate limits, per-call pricing, public listings, wallet balance, earnings, Stripe-backed payouts, environment variables, logs, and redeploys. The guidance assumes the dock0 CLI is available when command-line operations are needed.

japan-recs supports both single-area restaurant research and multi-neighborhood trip planning. It can import places you already saved, add restaurants and activities around them, and organize results as Markdown, CSV, or Google Sheets with useful details and links. Markdown and CSV work directly, while Sheets output requires the Google Workspace CLI and OAuth setup.