Recently updated open-source projects

bg-janitor targets macOS leftovers after AI tools close: orphaned children, sandbox-protected processes, port-holding gateways, and KeepAlive LaunchAgents. Its playbook distinguishes unsandboxed signal delivery, safe multi-PID handling in zsh, and launchctl unload -w for agents that immediately return after kill. Because the commands can terminate processes or disable startup services, confirm the target before applying them.

docs-to-md crawls an in-scope product documentation site and converts each page into an OKF v0.1 Concept: a Markdown file with YAML provenance and navigation metadata. It builds per-directory listings, a glossary, change log, and bundle index, rewrites internal links to relative Markdown paths, and validates the result against the OKF specification. The crawler runs with uv and keeps images as absolute source URLs instead of downloading them, giving agents a navigable local mirror.

Use this Claude skill to reshape AI-assisted drafts around a more natural voice, varied sentence rhythm, and clearer structure without intentionally changing the underlying facts. An optional Python scorer compares statistical detection signals before and after editing, but the project explicitly warns that detector results are unreliable and cannot guarantee passage against every trained classifier.

humanizer-fr rewrites pasted French text or a file to remove formulaic language-model habits, translated phrasing, and French typography mistakes. It adapts 33 patterns from the English humanizer and can calibrate the voice against a sample of the user’s own writing. The skill is prompt-only: it edits text but contains no executable code, making it suitable for Claude Code or OpenCode.

ziwei-bazi-reading calculates Zi Wei Dou Shu and BaZi charts with separate local engines, then injects LLM interpretations into an interactive HTML reading book with charts and yearly or major-cycle views. The charting can run offline and keeps interpretation JSON separate; the project frames astrology as cultural or entertainment material, not empirical science or professional advice.

Deep Discovery adds a Codex skill that interrogates a topic with 100 questions before you commit to a design, plan, or strategy. Each question builds on the previous answer, helping surface hidden assumptions, missing constraints, and weak decisions. It routes the process through focused patterns for architecture, code review, plugin and skill audits, product strategy, trading systems, or general analysis, making it useful as a deliberate pre-execution review step.

SuperICP turns a vague “we sell to everyone” position into a focused Ideal Customer Profile through a short, structured interview. It forces facts, guesses, buying context, and disqualifiers into the open, then returns an operational profile, lead scorecard, and validation plan for sales, advertising, and copy. It runs in any prompt-capable agent with no install or API key, across B2B, B2C, creator, and nonprofit contexts.

Frontcraft asks for a clear visual direction, palette, and level of expression before Claude Code writes the interface. It applies deliberate typography, color, layout, motion, and real imagery across common frontend stacks instead of defaulting to generic centered templates. After the build, review agents check issues such as headline collisions, dead links, stale images, and weak hierarchy; a Pexels API key is optional for more targeted photo search.

These agent skills help create and update common SAP S/4HANA records in public cloud, private cloud, and on-premise editions. They cover products, supplier invoices, purchase orders, business partners, service entry sheets, goods receipts, confirmations, communication setup, and a generic fallback, with CSRF handling, master-data lookup, idempotent batches, and transient retries. Users must configure tenant credentials, authentication, and the required communication arrangements before making changes.

Influencer Profile Board resolves which creator, expert, founder, or public personality you mean, gathers corroborated public signals, and renders one interactive HTML board with identity, contact channels, expertise, work, highlights, sources, and an as-of date. An optional SE Ranking connector adds live visibility data from ChatGPT and Google AI surfaces, but the standard web-search board works without it. Privacy rules exclude home addresses, private contact details, sensitive inferences, and gossip.

Product Modelling — Cloning an Influencer provides an eight-step method for turning a reference video into product-modeling stills and videos. It extracts frames, builds a character sheet, creates identity-anchored variants, adds products and storyboard shots, then makes a local fast-cut montage with frame-by-frame checks. The repository ships scripts and guidance only: no likeness, brand footage, or product photos are included. Users must supply licensed material and obtain the real person’s consent before cloning or publishing a likeness.

Provide a project name, description, author, and a few optional settings to generate a Python repository that is ready to initialize and push to GitHub. The Bash script or Claude Code skill creates the core documentation, packaging, tests, automation, and collaboration files, with its generated structure exercised by evaluations and GitHub Actions. It validates project scaffolding, not whether the underlying idea solves a real user problem.

Run the skill in a website repository to detect the stack, choose separate policies for retrieval and training crawlers, and audit the existing foundations. After approval, it can add or repair robots.txt, llms.txt, sitemap wiring, canonical and meta tags, and Schema.org markup, then verify the deployed files. These signals may improve discovery and attribution, but they do not guarantee AI citations, and robots.txt remains advisory.

client-secret-exposure audits web, mobile, and backend projects for secrets exposed in client bundles, insecure storage, broken access control, XSS, weak authentication, sensitive logs, dependency issues, headers, rate limits, business-logic flaws, and CI secrets. It is useful for Next.js and other app stacks, but the checks only identify findings; developers still need to remediate and retest the application.

clarify interrogates a request before work begins, identifying assumptions and presenting up to four structured, clickable questions with recommended options. It is useful when an incorrect implementation would cost more than a few extra messages, but it is manually invoked and never activates just because a prompt happens to match its description. The local plugin collects no data.

Spearmint helps developers create Jest tests by selecting a project and assembling arrangements, actions, and assertions through autocomplete, dropdowns, and guided forms. It can generate tests for accessibility, endpoints, GraphQL, Puppeteer, React, Vue, Svelte, Solid, Redux, hooks, and security-related workflows without requiring every line to be written manually. The result appears as editable test code. Export is currently limited to plain JavaScript; TypeScript test-file export remains planned rather than available.

TRAIL is a benchmark dataset with 148 annotated agent execution traces and 841 errors across reasoning, execution, and planning. The traces come from software-engineering and information-retrieval tasks, and the evaluation scripts run a LiteLLM-compatible model on GAIA or SWE Bench splits before calculating scores. It is intended for studying error localization in complex agent workflows.

Living Docs treats documentation as a governed part of the codebase, organizing constitutions, PRDs, ADRs, BDRs, issues, research notes, and Mermaid diagrams around five invariants. Facts have one home, every document is indexed, old decisions are superseded rather than rewritten, and structural changes carry their documentation in the same change. A deterministic linter checks the contract across Claude Code, Cursor, Copilot, OpenCode, Codex, and Pi. It governs organization, not a project’s technology choice.