Recently updated open-source projects

did-it-actually decomposes a user request into falsifiable acceptance criteria, records implementation claims, and audits the result with deterministic checks plus a fresh-context critic. Missing or unsupported claims trigger a fix-and-re-audit loop capped at three iterations, with a mechanically grounded VERIFIED or NOT VERIFIED verdict. It checks request fidelity rather than serving as a code reviewer, security scanner, or test framework.

This skill makes Claude inspect local files, system and Git history, authenticated MCPs, current web documentation, and delegated agents before giving up on a research-shaped question. A Stop hook blocks premature hedging unless enough tiers were tried or a structured escalation report is present. Its tier detection is heuristic and its phrase catalog mainly covers English and Portuguese.

landed maps each artifact to where users actually receive it: a remote branch and green CI, a live URL, an install path, a database row, a registry, a running process, or an email’s Sent folder. Its manual command and proactive Stop hook check that destination rather than trusting a local green build or successful push. The hook is intentionally narrow and blocks only strong ship claims with no destination evidence.

upshift defines five execution gears: inline edits, fan-out to parallel agents, a Workflow for many similar units, background work for one long job, and scheduled runs for recurring or polling tasks. It can be invoked manually or use a soft prompt hook that suggests the lowest gear fitting the work’s shape. The hint never blocks, and the one-gear-per-job rule helps avoid both under- and over-shifting.

flutter-rs provides tooling and a template for building desktop applications with Dart and Rust. After installing Rust, the Flutter SDK, and the cargo-flutter command, developers can create a project, run it with hot reloading, and build an AppImage release. The repository describes the project as work in progress, so it is better suited to experimentation, learning, or contributors willing to handle compatibility issues.

phinq-governance adds a procedural memory layer that an autonomous agent is expected to consult before taking a structurally significant action. It classifies reversible and irreversible work, checks the action against rule documents such as SOUL.md or AGENTS.md, pauses for confirmation when risk is high, and records events in an append-only tamper-evident log. It covers deletions, bulk operations, external communications, credential access, and payment changes, but it is an advisory protocol rather than a runtime interceptor or sandbox.

Render Word, PowerPoint, or PDF pages and audit them with a structured visual rubric covering layout, typography, color, content, and coherence. The skill reports findings rather than proposing fixes; scanned PDFs can be checked for layout but are not OCR-checked, and Office rendering needs LibreOffice plus an Anthropic API key for the visual audit.

write-human diagnoses predictable vocabulary, even sentence rhythm, hedging, and tidy parallel structure, then rewrites in three layers: lexicon, structure, and rhythm. It re-voices the draft with a concrete point of view, challenges what still sounds synthetic, and reads the result aloud. The skill changes surface style; it does not remove cryptographic watermarks or honestly promise a zero-AI detector score.

Skill Doctor scans project, global, and system skill roots to show what is installed, which definition wins, where names or triggers collide, and where risky patterns appear. It offers a local browser UI, CLI reports, and immutable scan history while binding to localhost by default. Static analysis works locally, with optional model-assisted analysis available through a configured compatible endpoint.

OysterWorkflow records screen states, OCR text, clicks, keystrokes, windows, and optional narration on macOS or Windows, then identifies meaningful work patterns and turns them into reviewable OpenClaw skill artifacts with retries, decisions, and completion checks. Users inspect the captured evidence and generated steps before installing anything into an agent. The app requires sensitive recording and accessibility permissions, currently supports Apple Silicon macOS and Windows x64, and has limited Chinese input and speech support on Windows.

Animated SVG Icons gives coding agents a design system and examples for interactive React SVG icons. It covers two-stage hover behavior, smooth return on hover-off, Framer Motion and requestAnimationFrame implementations, masking patterns, unique IDs, and cleanup. More than 20 TSX examples show how to adapt the system to a UI.

This skill and toolkit consolidate Scoreplan's v2 External, internal v2, and classic v1 API surfaces, including request schemas and a catalog of 109 endpoints. A Python utility can bulk-import actions, programs, projects, phases, and values from CSV or Excel, with a dry-run option for previewing requests. Creating indicators is not supported through the External API, and users must supply their own Scoreplan credentials.

Use one lead Claude Code session to spawn interactive workers in tmux, give each a task brief, coordinate through shared files, monitor progress, and require a different worker to challenge each result. The harness avoids servers and daemons but consumes parallel model usage and grants workers real filesystem and network permissions. Run it in a disposable or well-controlled environment.

The Craaft skill helps Claude Code create and manage Kanban cards through Craaft's JSON API. It documents bearer authentication, error recovery, rate limits, board-level access, and the full v1 endpoint surface, while calling out traps such as float positions, column keys, PATCH-only metadata, and multipart attachments. It also includes curl examples and guidance for the official Python SDK; an API token is required.

naturalize uses a SessionStart hook to nudge Claude toward natural prose throughout a session while skipping code, commits, and file paths. Each run rewrites the draft, rereads it as a skeptic, fixes passages that still sound synthetic, and ends by removing em dashes and en dashes. You can install it as an on-demand skill instead, and its test corpus covers the cataloged patterns.

Humanizer Pro can rewrite a draft, run a full editorial audit, or report AI-writing tells without changing the text. It checks promotional padding, formulaic rhythm, placeholders, citation residue, tracking links, and other artifacts, then applies restrained edits rather than forcing a casual voice. Its local CLI can also compare revisions for drift in numbers, dates, names, URLs, citations, quotes, and code blocks, with no external detector API.

design-room evaluates a website through seven codified lenses, covering first impression, conversion usability, decision psychology, positioning, technical accuracy, and feasibility or boldness. In grounded mode Claude assembles source-backed options while you decide; a weekly modernity pulse and anti-slop gate challenge stale or generic output. Each decision is appended to a dated, hash-chained decisions.md record so later sessions can see what changed and why.

AgentTier runs isolated, persistent sandboxes for AI agents and coding assistants on Kubernetes. Each sandbox keeps its workspace on a PVC, applies network and resource policies, and exposes an authenticated terminal, file, command, and agent-invoke surface. The platform includes lifecycle management, templates, governance, audit events, OIDC, and optional gVisor isolation, and requires Kubernetes 1.27+, Helm 3, and compatible NetworkPolicy and CSI storage.