Recently updated open-source projects

academic-topic-research-agent turns a one-sentence research idea into a preliminary package for venues such as CHI, CSCW, LAK, and AIED. Quick, standard, and deep modes can produce literature metadata, feasibility verdicts, title and abstract options, research questions, method boundaries, evaluation plans, risks, next steps, venue positioning, and reviewer-attack predictions. A concept stress test adds multi-angle critique and revision; it is not a paper-writing service or a substitute for scholarly judgment.

Urd activates when an agent moves from a plan or design into implementation and encounters an uncovered edge case, contradiction, unsanctioned fallback, weakened test, or uncertain workaround. It prompts the agent to explain the exact step, quote the relevant specification, describe the gap, present options, and recommend a choice while separating verified facts from beliefs. This is a behavioral guardrail rather than a runtime guarantee, and it deliberately trades extra clarification rounds for fewer silent deviations.

figure-extractor finds complete figures and images in PDFs, arXiv papers, or HTML, using captions and bounding boxes to avoid cutting off multi-panel content. It crops with PyMuPDF at a chosen DPI and can build a contact sheet for quick visual checking. The project is a shell CLI, not a hosted or in-agent service, so Python, PyMuPDF, and a pip-capable runtime are required.

UI/UX Consultant Skill gives Claude a multi-framework design reference for interface requests. It covers 16 frameworks with opinionated code patterns, accessibility checklists, 23 visual styles, 35 color palettes, 31 font pairings, and layout guidance for 41 product types. It can be initialized for Claude Code or Cursor and is intended for design direction plus implementation patterns rather than a visual editor.

claude-code-deepseek-3p-parity-skill inspects Claude and MCP settings, permissions, proxies, and runtime versions in dry-run mode, then proposes configuration changes for DeepSeek or other compatible gateways. After confirmation it can add model slots, permission rules, trusted workspaces, and MCP entries while backing up JSON and redacting secrets. It cannot grant first-party model capabilities, bypass macOS security prompts, or fix provider limitations.

timplay moves product planning through four dependent stages: a question-driven PRD, an editable feature tree, a user-flow diagram, and a clickable wireframe prototype. The HTML tools support drag-and-drop, node and connection editing, responsive desktop/mobile views, and JSON round-trips for browser edits. Generated product content follows the input language, but the surrounding HTML controls are currently fixed in Korean.

logfrisk checks logging code and representative output for bearer tokens, emails, password hashes, error payloads, and other secrets or personal data. Findings include file and line locations, severity, and a suggested redaction; clear leaks can be replaced with redacted fields, while ambiguous cases are escalated for a decision. The skill keeps the work open until every leak is fixed or explicitly resolved.

Humanize detects 29 concrete patterns associated with machine-written prose, then iterates through transformation, re-scoring, and keep-or-discard decisions for up to ten rounds. A writing sample lets it match your rhythm, vocabulary, and punctuation instead of imposing a generic voice. The final pass removes em dashes, en dashes, and double hyphens, and the plain skill runs without code dependencies.

The skill detects architecture, security, performance, debugging, code-quality, and context-management situations where a stronger model may help. It extracts a focused problem statement, relevant code, failed approaches, and constraints, then invokes an advisor through the coding platform’s native model or sub-agent mechanism; multiple domain advisors can be consulted in parallel. Its operation therefore depends on the selected tool actually supporting model switching or agent dispatch and having the requested advisor model available.

safedeps grounds dependency choices in current OSV vulnerability records instead of a model’s frozen training knowledge. When an agent is about to add a package and version, it checks for known flaws and can stop the suggestion before it enters the project. This is a narrow decision-time guard, not a replacement for npm audit, pip-audit, osv-scanner, or a broader CI security scan.

Goal by Goal turns a plan, PRD, or brief into 5–12 sequenced goals, with a brief, tracker, reviewer prompt, and check script for each. Build, test, or lint checks are required, and each goal must pass its read-only check before a commit.

Petros Methodology gives Claude and Claude Code a set of taste and workflow rules for writing, documents, and interface work. It emphasizes concise language, restrained punctuation, dense layouts, autonomous execution, and refusing to invent missing data. Before declaring a task complete, it asks the agent to verify the real result through checks such as end-to-end flows, mobile behavior, or a confirmed deployment.

html-animation-to-mp4 captures HTML/CSS/JS animations in a headless browser and encodes the frames with ffmpeg. Claude Design .dc.html bundles can expose their own frame-seek hook, while --virtual freezes clocks and advances ordinary canvas, WebGL, or CSS animation frame by frame. It supports multiple aspect ratios and frame rates; capture is silent by default, and Playwright, Node, and ffmpeg are required for the CLI.

db-route-optimizer identifies N+1 queries, missing indexes, queries in loops, unbounded reads, and other route bottlenecks through query counts or EXPLAIN ANALYZE rather than guesswork. It saves a baseline, applies one isolated behavior-preserving change, re-benchmarks under the same conditions, and records the result. Tests and a query-count guard can stay in CI to prevent regressions, and the scripts also run without Claude.

OpenBioHack helps a person reason through difficult health questions using their own data, keeping several possible mechanisms open instead of collapsing to a diagnosis. It can rank low-risk, reversible options to consider and observe, while showing the reasoning so the person can discuss it with a qualified clinician. The project is educational and non-directive: it is not medical advice, a diagnosis, or a medical device, and it does not make decisions for the user or clinician.

Glucose Media Player plays local video and audio files on Windows while generating subtitles with a local Whisper model. It can load common subtitle files, extract embedded tracks, organize selected folders into a gallery, and keep playback visible through picture-in-picture. The app is currently Windows-only and is aimed at people who want private transcription without uploading media or relying on an internet connection.

revisao-literatura-medica is for medical professionals and supports literature reviews across clinical topics, with POCUS as one use case. It searches PubMed, Europe PMC, and OpenAlex, verifies PMIDs and DOIs, resolves open-access copies through Unpaywall, and reads PDF or EPUB documents TOC-first, using OCR for scans. Outputs follow evidence-grading and guideline or study structures, and the skill only states claims retrieved from a real source in the current session. It requires Python dependencies; Telegram-library access stays in private configuration.

design-effects-skill teaches an agent to select visual effects for UI and branding contexts, with decision rules that also say when not to use them. Its references cover animated backgrounds, text and scroll motion, liquid or glass treatments, shader gradients, and 3D UI, with implementation examples tied to common libraries. A browser-ready examples page lets you preview the effects without a build step.