GitHub Trending Weekly Digest — Aug 24–29, 2026

If you blinked this week, you missed a leaderboard reshuffle. Between August 24 and August 29, one prompt library camped at the top for four straight days, an AI diagramming tool clawed its way from newcomer to number one, and the whole "agent skills" movement went from niche to unmistakable.

I pulled the daily GitHub Trending Top 5 for all six days, merged them, removed duplicates by repository URL, and ranked everything by how many days it actually stuck around. Nineteen distinct projects made the cut. Here's the week, sorted by staying power.

🏆 Persistent Chart-Toppers

These two didn't just show up — they refused to leave.

awesome-gpt-image-2 — 4 days on the board

🔗 github.com/freestylefly/awesome-gpt-image-2

What it does: An industrial-grade prompt engine for GPT image generation, packing 530+ reverse-engineered prompts into 20+ reusable production templates across a dozen categories — UI design, infographics, photography, illustration, and more.

Why it matters: AI image generation has moved past "can it draw this?" to "can it draw this reliably, controllably, and repeatably?" The community's best prompts have always been scattered, one-off, and hard to reuse. This project treats "Prompt as Code" — it breaks long prose prompts into atomic, composable modules (subject, lighting, material, layout, visual detail) so you can batch-generate and automate the whole workflow. It even ships as an agent skill package you can plug into Claude Code, Cursor, or Codex.

Tech: Vite + React/Vue frontend, Supabase (Auth + PostgreSQL), Vercel Functions as the API proxy to GPT-Image2, Stripe and Alipay for billing, plus an npm-published agent skill with structured JSON export.

archify — 3 days, and it finished the week at #1

🔗 github.com/tt-a1i/archify

What it does: An AI-agent-driven Node.js tool that turns a codebase or a system description into interactive, verifiable architecture diagrams — flowcharts, sequence diagrams, data-flow and lifecycle views included.

Why it matters: Hand-drawn architecture docs rot. They drift away from the real code and become impossible to trust. Archify compiles diagrams deterministically — same input, same output — so charts are generated only from facts the author actually wrote, never hallucinated topology. You can diff architecture snapshots before merging, and nodes bind back to specific files in specific Git commits, giving reviews a paper trail with machine-validated receipts. Its climb over the week — roughly a thousand stars a day, then four thousand-plus on the final two days — tells you developers were hungry for exactly this.

Tech: Node.js rendering engine with a typed JSON intermediate representation, a built-in schema validator, self-contained HTML/SVG output, and PNG/SVG/WebM export. Works as an agent skill inside Cursor, Claude Code, and Codex CLI.

📈 Multi-Day Regulars

Two-day residents — enough momentum to prove the first day wasn't a fluke.

Codex — OpenAI's terminal coding agent

🔗 github.com/openai/codex

What it does: A lightweight coding agent from OpenAI that runs right in your local terminal, with support for VS Code, Cursor, Windsurf, and a desktop app.

Why it matters: It lets you write code and run tasks with AI without leaving your workflow for a browser chat window. Everything runs on your own machine, and you can reuse an existing ChatGPT subscription or bring your own API key — privacy, control, and cost all handled. It opened the week strong with a combined 4,700 stars across two days.

Tech: Core written in Rust (codex-rs) with a TypeScript/JavaScript CLI, built with Bazel, Nix flake support, and prebuilt binaries for macOS and Linux. Apache-2.0.

God's Eye View — a spy satellite sim with real data

🔗 github.com/bilawalsidhu/gods-eye-view

What it does: A browser-based geospatial intelligence platform that visualizes live global data — aircraft, ships, satellites, earthquakes, traffic, public cameras — on a photorealistic 3D Earth. The author's pitch: "a spy satellite simulator in your browser, except the data is real."

Why it matters: As the creator puts it, most open-source intelligence work is just "a pile of browser tabs." God's Eye View folds a dozen-plus public data sources into one intuitive 3D interface, with thermal and night-vision sensor modes, AI voice commands, and shareable cinematic camera tours. Scattered signals become an explorable global situational map.

Tech: Vanilla JavaScript (deliberately framework-free for readability) + CesiumJS for 3D, Vite for builds, Google Photorealistic 3D Tiles for the Earth, and the OpenAI Realtime API driving 28 voice-control tools. Data comes from OpenSky, AISStream, CelesTrak, USGS and other mostly-unauthenticated public feeds.

ai-job-search — a job hunt that actually landed one

🔗 github.com/MadsLorentzen/ai-job-search

What it does: A Claude Code–based job search framework that runs locally to score job fit, tailor résumés and cover letters, and prep you for interviews.

Why it matters: Job hunting is repetitive and draining. This tool runs a drafter-reviewer dual-agent workflow — one agent writes the materials, another critiques them — then compiles a PDF and checks it for ATS (applicant tracking system) readability. It leans hard on being deterministic and verifiable: it won't invent credentials, and every claim is checked against your real profile. The kicker: after losing a geophysics job, the author used it to send 69 applications, landed 20 first-round interviews, and pivoted into an AI engineering role.

Tech: Claude Code as the reasoning core, Python 3.10+ for orchestration, LaTeX (moderncv) for document generation, poppler for PDF text extraction, and a Bun-based CLI hitting LinkedIn plus several Danish job portals. Four commands: /setup, /scrape, /apply, /interview.

claude-plugins-official — Anthropic's curated marketplace

🔗 github.com/anthropics/claude-plugins-official

What it does: Anthropic's officially maintained, curated directory of high-quality Claude Code plugins — vetted official internal plugins alongside third-party community ones.

Why it matters: It gives developers a single trusted place to discover and install plugins that extend Claude Code, held to consistent quality and security standards so you don't get burned by sketchy extensions. A two-tier layout (/plugins for official, /external_plugins for third-party) keeps the sources clear, and a single /plugin install {name}@claude-plugins-official gets you going.

Tech: Built on the Claude Code plugin system with MCP (Model Context Protocol) server support, slash commands, and agent definitions. Metadata lives in plugin.json with an optional .mcp.json; distribution is Git-based.

go-modern-guidelines — teaching agents to write current Go

🔗 github.com/JetBrains/go-modern-guidelines

What it does: A set of guidelines and tooling that helps AI coding agents write "modern Go" — using max(a, b) instead of an if-block, slices.Contains for list checks, and the newer standard library from Go 1.26–1.27.

Why it matters: Coding agents tend to write dated Go for two reasons: training data lags behind new language features, and older idioms simply appear more often in the corpus. This guide gives agents an explicit reference so they write modern Go from the first draft, cutting down on later refactoring.

Tech: A Go CLI tool distributed via go install, auto-detecting your project's Go version from go.mod (targeting 1.25+), and plugged into Claude Code, Cursor, Junie, and Codex through their plugin marketplaces.

nitter — the privacy front-end that got a C&D

🔗 github.com/zedeus/nitter

What it does: A free, open-source, JavaScript-free alternative front-end for Twitter, built around privacy and speed.

Why it matters: Twitter's own pages track you through JS telemetry, browser fingerprinting, and IP logging with no way to opt out. Nitter proxies every request through its own backend so the client never touches Twitter directly, cutting tracking off entirely — and it's roughly 15 times lighter than the official site. Worth noting: the repo was archived in August 2026 after a cease-and-desist from X Corp., so it's now read-only. Its two quiet days on the chart read like a farewell wave.

Tech: Written in Nim, with Redis/Valkey for caching, libsass for styles, and libpcre for pattern matching. Deploys via Docker, systemd, or the traditional route, usually behind an Nginx or Apache reverse proxy.

🌱 One-Day Wonders

They flashed onto the board for a single day — but plenty of these are worth a bookmark.

mattpocock/skills — engineer-grade agent skills

🔗 github.com/mattpocock/skills

What it does: A set of AI agent skills Matt Pocock pulled straight out of his own .agents directory, meant "for real engineers" and made to work with Claude Code, Codex, and friends.

Why it matters: It targets the four classic failure modes of AI-assisted development: going off the rails (building something you didn't ask for), rambling output, low code quality (won't run, or full of bugs), and architectural rot (the codebase turning into a "ball of mud"). Structured workflows tackle each one.

Tech: Mostly Shell, orchestrating work through user-triggered commands like /grill-me, /tdd, and /to-spec, grounded in DDD, TDD, and ADR methodology. Install via the Claude Code marketplace or npx skills@latest add mattpocock/skills.

OpenLogi — a local-first Logitech Options+ replacement

🔗 github.com/AprilNEA/OpenLogi

What it does: A native, local-first replacement for Logitech Options+, written in Rust, for managing your Logitech peripherals.

Why it matters: The official Options+ demands an account, ships telemetry, and treats Linux as a second-class citizen. OpenLogi runs fully local with no account and no data collection, uses plain-text TOML config, supports custom gestures and CLI scripting on any key, and makes Linux a first-class platform across macOS, Linux, and Windows.

Tech: Rust + GPUI, driving mice and keyboards over HID++ (DPI, key remapping, SmartShift, RGB) and webcams over UVC. Background device I/O is split from the GUI, and every setting lives in a single portable TOML file.

free-claude-code — one gateway, many agents

🔗 github.com/Alishahryar1/free-claude-code

What it does: An aggregator platform that calls Claude Code, Codex, Pi, OpenCode, Cline and other AI coding agents through one unified entry point, claiming access to over 1.3 billion free tokens a month.

Why it matters: It tackles model fragmentation and cost head-on — aggregating 50+ free and paid model providers, auto-falling-back when one goes down so you keep coding through outages, and filtering terminal output to cut token usage by up to 90 percent.

Tech: Python 3.14 with uv, pytest, Ruff, and Loguru. Connects to NVIDIA NIM, OpenRouter, Groq, OpenAI, Gemini, DeepSeek, Mistral, Ollama, LM Studio, and llama.cpp, with clients spanning CLI, VS Code, JetBrains, Discord/Telegram bots, and mobile — plus local Whisper voice transcription.

omarchy — DHH's opinionated Linux distro

🔗 github.com/basecamp/omarchy

What it does: A "beautiful, modern, opinionated" Linux distribution built by DHH (Basecamp's founder).

Why it matters: Traditional Linux needs a lot of tinkering before it's pleasant. Omarchy ships an out-of-the-box, pre-configured, polished desktop, aimed squarely at people coming from macOS or Windows who just want a good-looking environment that already works.

Tech: Shell scripts + dotfiles for config management, with a full desktop and theming system, a dev toolchain (Neovim, terminal, shell tools), AI integration, and browser and gaming support. A 51-chapter manual, MIT-licensed.

scientific-agent-skills — 163 skills for research agents

🔗 github.com/K-Dense-AI/scientific-agent-skills

What it does: An open skill library of 163 ready-to-use research skills that turn any AI agent into a research assistant capable of running complex scientific workflows across bioinformatics, drug discovery, clinical research, and dozens of other fields.

Why it matters: Agents can technically call any Python package or API, but what they lack is domain-specific guidance. Instead of letting an agent fumble through docs, this library hands it explicit, version-aware workflows with examples, best practices, and safety boundaries — making it far more reliable on scientific tasks.

Tech: Python 3.13+ with uv, following the open Agent Skills spec. Integrates 100+ scientific databases (PubChem, ChEMBL, UniProt, ClinicalTrials.gov) and 70+ optimized Python-package skills (RDKit, Scanpy, PyTorch Lightning). Works with Cursor, Claude Code, and Codex. MIT.

andrej-karpathy-skills — Karpathy's coding-agent guardrails

🔗 github.com/multica-ai/andrej-karpathy-skills

What it does: A CLAUDE.md file (plus Cursor rules) distilled from Andrej Karpathy's observations on the common pitfalls of LLM-written code, meant to improve how Claude Code and Cursor behave.

Why it matters: It targets the usual LLM bad habits — making unwarranted assumptions, over-complicating solutions, editing unrelated code, and working without clear success criteria. Four core principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution.

Tech: No special stack — it's a Markdown instruction set. Install via the Claude Code marketplace, drop it into a project, or use it as a Cursor rule. MIT.

maka — a local-first agent workspace (Apache incubating)

🔗 github.com/apache/maka

What it does: A local-first AI agent workspace that records every model interaction, tool call, and result to an append-only log on your own machine.

Why it matters: It's built around privacy and vendor-lock-in worries. Sessions and records stay local by default, and you bring your own model — cloud API, local model, or gateway — while keeping a complete, recoverable execution history that doesn't depend on the cloud.

Tech: Node.js + TypeScript, an Electron + React desktop app, SQLite storage, with CLI/TUI support. Architecture splits into Runtime Host, SessionManager, and AgentRun layers. Currently incubating at Apache, with an early public build for macOS Apple Silicon.

claude-plugins-community — the vetted mirror

🔗 github.com/anthropics/claude-plugins-community

What it does: Anthropic's officially maintained read-only mirror of the community plugin marketplace for Claude Cowork and Claude Code.

Why it matters: It gives users one central, reviewed place to discover and install extensions instead of hunting across repos. Every plugin passes automated security scanning and syncs nightly from Anthropic's internal review pipeline; directly-filed PRs are auto-closed, so all changes go through internal review.

Tech: Built on the Claude Agent SDK, with marketplace data in .claude-plugin/marketplace.json. Install via claude plugin marketplace add anthropics/claude-plugins-community.

plane — open-source project management

🔗 github.com/makeplane/plane

What it does: An open-source project management platform positioned as an alternative to Jira, Linear, Monday, and ClickUp — for tracking issues, running sprint cycles, and managing product roadmaps.

Why it matters: Teams want a flexible, self-hostable PM tool without getting locked into a proprietary ecosystem. Plane gives you data ownership and transparency, with Work Items, Cycles (including burndown charts), Modules, custom views, AI-enabled collaborative Pages, and real-time analytics.

Tech: React Router + Vite frontend, Django + Node.js backend, PostgreSQL storage, Redis caching, with Docker and Kubernetes deployment. AGPL-3.0.

TradingAgents — a simulated trading firm

🔗 github.com/TauricResearch/TradingAgents

What it does: A multi-agent financial trading framework that uses several LLM-driven specialist agents to simulate how a real trading firm makes decisions.

Why it matters: It splits complex trading analysis into distinct roles — fundamental analyst, sentiment expert, technical analyst, trader, risk manager — that collaborate on a structured evaluation of fundamentals, market sentiment, technical patterns, and risk, instead of leaning on one monolithic "do-everything" analysis.

Tech: LangGraph for orchestration, Python 3.12, Docker, and a CLI. Supports OpenAI, Gemini, Claude, Grok, DeepSeek, Qwen, and local Ollama; pulls data from Yahoo Finance, Alpha Vantage, Reddit, and StockTwits; uses SQLite checkpoints for crash recovery and supports backtesting.

GitNexus — a knowledge graph for your codebase

🔗 github.com/abhigyanpatwari/GitNexus

What it does: A code intelligence engine that builds a knowledge graph of your codebase client-side and exposes it through AI agent tools — as a CLI local indexer, an MCP server for editors, or a browser UI for exploration.

Why it matters: AI coding assistants like Cursor and Claude Code lack deep understanding of code architecture, which leads to broken dependencies and half-finished refactors. GitNexus precomputes the structure at index time — clustering, tracing, scoring — so a single call returns full context instead of endless back-and-forth graph queries.

Tech: Tree-sitter for parsing (native bindings in the CLI, WASM in the browser), LadybugDB as an embedded vector-capable graph database, hybrid retrieval (BM25 + semantic embeddings + reciprocal rank fusion), and Sigma.js + Graphology for WebGL visualization. React 18 + TypeScript + Vite + Tailwind v4, with 17 MCP tools and support for a dozen-plus languages.

This Week's Themes

A few clear patterns ran through the six days.

Agent skills went mainstream

The single loudest signal of the week. "Skill packages" for coding agents were everywhere — and not as a novelty, but as the format serious people are shipping in:

awesome-gpt-image-2 ships prompts as an agent skill
mattpocock/skills packages engineering methodology as skills
andrej-karpathy-skills distills LLM guardrails into a skill file
scientific-agent-skills offers 163 research skills
go-modern-guidelines teaches agents modern Go through a plugin
archify itself installs as an agent skill

The Claude Code plugin ecosystem is consolidating

Two Anthropic-run marketplaces trended in the same week — claude-plugins-official (the curated directory) and claude-plugins-community (the vetted read-only mirror) — alongside free-claude-code, an aggregator routing across agents. The plumbing around coding agents is professionalizing fast.

Local-first and privacy stayed loud

OpenLogi — no account, no telemetry, pure local peripheral control
maka — an append-only local agent workspace
nitter — privacy-first, right up until the cease-and-desist
omarchy — an opinionated, self-owned desktop

Making the invisible visible

archify — codebases into verifiable architecture diagrams
gods-eye-view — scattered OSINT signals into a live 3D Earth
GitNexus — code into a queryable knowledge graph

Takeaway

If you take one thing from this week, make it this: the agent-skills format has crossed from experiment to expectation. Six of the nineteen projects are skills packages of one kind or another, and the two that dominated the chart — awesome-gpt-image-2 and archify — both ship as skills you drop straight into Claude Code, Cursor, or Codex. The question is no longer whether your tool has an AI integration; it's whether it ships as a skill the agent can pick up on its own.

Meanwhile, the classics still resonate. A privacy front-end got archived under legal pressure and people rallied to it anyway. A job-search framework trended because it worked — 69 applications, 20 interviews, one career pivot. And a solo-built 3D Earth reminded everyone that "vanilla JavaScript, no framework" can still stop the scroll.

See you next week.

Compiled by Tommy Zhang | August 30, 2026