Recently updated open-source projects

Snapshot Reader lets an agent work from two Markdown files: one file tree and one snapshot containing the source in code blocks. It extracts only the files needed for a request, tracks multiple snapshots across turns, and returns minimal BEFORE/AFTER fragments ready to apply in an editor. The workflow is useful when you do not want to grant direct repository access, but changes are applied by the user rather than written automatically.

Drop a prequel quote, character reference, or related phrase into a Claude Code conversation and this skill responds in matching Star Wars meme character without abandoning the requested task. It includes more than 120 call-and-response triggers across four tiers, making it a playful tone layer rather than a separate coding or analysis tool.

CleanMyMac and CleanMyPC inspect actual disk usage before proposing targeted cleanup. They separate caches from user data, ask for confirmation before each deletion, use Trash-first handling for larger or riskier items, and compare free space before and after. The skills cover macOS and Windows, but explicitly protect passwords, mail, cloud-synced files, browser profiles, and other sensitive locations. They are for negotiated cleanup, not blind sweeping.

Local Delegate decides when a large, repetitive coding task is worth sending to a local LM Studio model. It prepares a self-contained request, calls the local OpenAI-compatible endpoint, verifies the result, and falls back to direct writing when the server is unavailable. The skill targets outputs of roughly 80 lines or more, such as fixtures, enums, and test cases; it is not intended for refactoring, security-sensitive code, or small snippets. Its examples assume Windows PowerShell.

Use AIVC DACH to turn a short request into local MP4 templates for intros, shorts, sponsor reads, and similar formats; the renderer works without an account or API key. Its v3 alpha editor already supports chat-driven cuts, trims, and transparent overlays, while direct timeline control from external MCP agents is still planned. The toolkit requires Claude Code, Node.js, and Bun.

Article Analyzer deconstructs an article instead of declaring it good or bad. It classifies sentences, checks factual claims against primary sources, surfaces hidden premises and skipped logical steps, and tags rhetorical framing. Its final audit-style opinion is limited to factual claims, with no-opinion boundaries for political, religious, illegal, and certain entertainment content. It can report that a source was not found without treating that absence as proof of falsehood.

claude-roundtable turns a real question into a structured debate among installed perspective skills. It anchors the discussion in explicit decisions, variables, and assumptions, requires speakers to challenge weaknesses and propose alternatives, and ends with an integrated plan. Users can add a participant, update constraints, or jump to synthesis while the debate is running; at least one persona skill is required.

st-pricebook documents ServiceTitan Pricebook v2 schemas, endpoints, fields that silently disappear on writes, bulk Excel or CSV imports, configurable equipment, estimate templates, and static versus dynamic pricing. Its update pattern combines dry runs, HMAC confirmation, and read-after-write hash checks, with an incident log for recurring mistakes. The skill is tenant-agnostic, so users must replace the supplied tenant and sync placeholders before use.

Preflight sends a plan, specification, or RFC to three to five independent domain experts before implementation begins. It combines their findings into a ranked report with a verdict, must-fix items, explicit decision cards, cross-confirmed concerns, and gaps no panelist covered. A human gate confirms the proposed roster first; the workflow is intended for complex multi-domain designs, not trivial plans, debugging, or review of code that already exists.

roundtable-personas is a persona library for claude-roundtable. Each subdirectory is an independently installable perspective skill that the roundtable can discover automatically. The repository currently provides a Feynman perspective focused on scientific reasoning and logical verification, and documents the shared SKILL.md shape for adding more personas.

distill-ui reverse-engineers a codebase into a design system that people and coding agents can reuse, then builds an interactive showcase to prove the extraction matches the product. It produces design docs, canonical tokens, a component registry, and agent instructions, with optional Tailwind or React output, and can flag common AI-design patterns. It works from source code rather than guessing from screenshots.

Competitive Intel to Conversion Site turns real search results, competitor pages, and buyer reviews into a conversion-focused site strategy. It captures positioning, pricing, trust signals, and customer language, then ties each proposed section to a buyer job, objection, and proof source. The handoff includes an HTML report, site structure, wireframe, and copy pack; WebSearch/WebFetch or Firecrawl can automate collection, but manual use needs no API key.

Lingshu makes a primary agent the control plane for multi-model work. It defines scope and acceptance criteria before delegating to models, CLIs, scripts, or connectors, then reviews their increments and performs final acceptance. Phase reports, plus treating existing tests as protected contracts, help prevent drift, unauthorized actions, and weakening tests just to make a run pass.

Publish Skill turns the current chat's problem, steps, requirements, and anti-patterns into a reusable package: a local SKILL.md, a public GitHub repository with a bilingual README, and a GitHub Pages landing page. It can also generate a cover and add a card to a personal site; raster output is used in Codex with SVG fallback when image generation is unavailable. It requires authenticated gh, git, and bash; a site-repository path is only needed for the optional card.

Scroll-Scrub Hero guides Claude Code, Codex, or Gemini from two generated images to a roughly seven-second video and a scroll-controlled hero. The result lets users reveal an assembly or transformation as they scroll, with staged tooltips and mobile autoplay or reduced-motion fallbacks. It supports Next.js 16+ or vanilla HTML and requires image and video service accounts.

Start Goal converts a vague request into a measurable goal brief before autonomous work begins. It restates assumptions, classifies the task, gathers the needed context and constraints, defines deliverables and stop conditions, and writes a concrete “Done when” checklist. Creative or product tasks pass through brainstorming first, and long-running work is handed to /goal only after acceptance criteria are specific.

kaizen-consulting-skill helps consultants plan organizational or tool changes as stages instead of a single Big Bang rollout. From a project brief, it sets transition criteria around roughly 70% acceptance, segments employees into four response groups, and proposes communication, risks, anti-patterns, and a next step with an owner and deadline. Its trigger phrases are primarily German, so English activation may be unreliable.

Session Handoff packages a Claude session into a structured file and copies a ready-to-paste opening message for the next chat. It detects a development context and creates a compact, plan-linked pointer, or switches to a fuller research handoff with findings, decisions, rejected alternatives, and open questions. Clipboard support uses pbcopy on macOS, with xclip or xsel as Linux alternatives.