GitHub Open Source Weekly 2026-02-25: Skills Ecosystem Solidifies, Embedded AI Rises, OpenClaw Offspring Sweeps Prediction Markets

Data period: 2026-02-18 – 2026-02-25 (rolling 7 days) Sources: GitHub Trending weekly + monthly, GitHub Search API, HN Algolia

TL;DR: The biggest surprise this week is the New Repos chart being flooded by prediction market tools spawned from the OpenClaw ecosystem—several of which carry serious security risks (details below). The weekly star-gain champion x1xhlol/system-prompts again confirms developers' unrelenting curiosity about AI tool internals. The sustained momentum signal comes from obra/superpowers, which added nearly 7,000 stars in a single week while staying on the monthly chart—marking the moment the Skills ecosystem formally graduated from personal experiment to framework infrastructure.

📈 Fastest Growing — Weekly Star Gains Top 10

Source: github.com/trending?since=weekly 🔁 = Also on the monthly trending list (sustained momentum signal)

#	Repo	+Stars/week	Total Stars	Language	Created
1	x1xhlol/system-prompts-and-models-of-ai-tools	+7,784	123,703	—	2025-03
2 🔁	obra/superpowers	+6,964	61,201	Shell	2025-10
3	alibaba/zvec	+3,460	7,839	C++	2025-12
4	huggingface/skills	+3,381	6,117	Python	2025-11
5	anthropics/claude-code	+2,414	70,004	Shell	2025-02
6	google-research/timesfm	+1,903	9,725	Python	2024-04
7	Stremio/stremio-web	+1,087	10,104	JavaScript	2018-06
8	muratcankoylan/Agent-Skills-for-Context-Engineering	+1,072	10,418	Python	2025-12
9	cloudflare/agents	+940	4,215	TypeScript	2025-01
10	SynkraAI/aios-core	+707	1,805	JavaScript	2025-12

🆕 Top New Repos — This Week's Newcomers Top 10

Source: GitHub Search API (created:2026-02-18..2026-02-25, sorted by total stars) ⚠️ = Abnormal stars/forks ratio — possible star inflation or malware risk

#	Repo	Total Stars	Language	Created
1	cloudflare/vinext	2,172	TypeScript	2026-02-24
2	Leonxlnx/taste-skill	1,524	—	2026-02-19
3	ShinMegamiBoson/OpenPlanter	1,310	Python	2026-02-20
4	anthropics/financial-services-plugins	905	Python	2026-02-23
5	RightNow-AI/picolm	882	C	2026-02-19
6	olvvier/apple-silicon-accelerometer	797	Python	2026-02-19
7	Polymarket/polymarket-cli	770	Rust	2026-02-24
8	Panniantong/Agent-Reach	731	Python	2026-02-24
9 ⚠️	Kirubel125/Kalshi-Claw	690	TypeScript	2026-02-22
10 ⚠️	CraftyGeezer/Kalshi-Polymarket-Ai-bot	680	Python	2026-02-21

Spotlight — Fastest Growing Top 10

📈 #1 — x1xhlol/system-prompts-and-models-of-ai-tools｜The Ultimate AI Tool System Prompt Collection

FULL Augment Code, Claude Code, Cluely, Cursor, Devin AI, Lovable, Manus, Perplexity, Replit, Windsurf, v0... System Prompts, Internal Tools & AI Models

+7,784 ★ this week｜123,703 total｜GPL-3.0

The premise is simple: collect the system prompts of every major AI coding tool (Cursor, Claude Code, Windsurf, Devin, v0, and more) so anyone can see what instructions are actually running inside these black boxes. Nearly 8,000 stars in a week. There is one related HN thread this week—low on points, but 60,000+ forks signal that people are actively pulling this apart to study it.

What this means for developers: you can learn directly from how top AI tools design context windows and constrain model behavior—a shortcut to better system prompt engineering for your own AI applications.

📈 #2 🔁 — obra/superpowers｜The Pioneer of the Skills Era

An agentic skills framework & software development methodology that works.

+6,964 ★ this week｜61,201 total｜Shell｜MIT

obra is Jesse Vincent—co-founder of Keyboardio (ergonomic keyboards), founder of Best Practical (Request Tracker), former Perl pumpking. He released superpowers in October 2025: a composable "skills" framework designed for Claude Code.

The core idea: break your development workflow into individual markdown instruction files (TDD protocol, debug methodology, subagent delegation patterns). When the AI receives a task, it steps back to clarify requirements, produces a spec, then launches subagents to execute in parallel.

+6,964 stars this week while staying on the monthly chart (🔁)—the only monthly holdover this week. Two months of sustained growth means real production usage, not hype.

📈 #3 — alibaba/zvec｜The SQLite of Vector Databases

A lightweight, lightning-fast, in-process vector database

+3,460 ★ this week｜7,839 total｜C++｜Apache-2.0

Alibaba's open-source embedded vector database runs directly inside your application process—no separate server, no Docker. The HN 225-point discussion was the week's highest-temperature technical debate.

Technical highlights:

Built on Proxima, Alibaba's internal production vector search engine
Claims >8,000 QPS on VectorDBBench, allegedly 5× OpenSearch and 19× Milvus
Supports dense + sparse hybrid search and multi-vector queries
Python and Node.js support

Two core HN controversies: First, self-reported benchmarks with no third-party verification—one tester found latency jumped from 0.8ms to 100ms+ after switching to cloud object storage (blobfuse2), severely limiting cloud-native viability. Second, no comparisons against DuckDB vector extensions, pgvector, or FAISS—Alibaba acknowledged this gap.

Community consensus: excellent as an embedded vector library for local RAG and edge deployments; not the right tool for distributed cloud architectures. The "SQLite of vector DBs" framing is accurate.

📈 #4 — huggingface/skills｜HuggingFace's Official Skills Repository

(No official description — inferred: an AI coding agent skill library)

+3,381 ★ this week｜6,117 total｜Python｜Apache-2.0

HuggingFace's official skills repository, up +3,381 stars alongside obra/superpowers and muratcankoylan/Agent-Skills—forming a clear signal: the Skills ecosystem formally shifted from individual experiments to platform support this week.

Worth noting: HN records from January 19 show someone already attempted a "NPM/uv for Claude Code" Show HN, indicating the community has been thinking about a central registry with package-manager-style installation. HuggingFace entering the space means the most influential ML platform is now building that infrastructure.

📈 #5 — anthropics/claude-code｜+2,414 Stars at Baseline

Claude Code is an agentic coding tool that lives in your terminal...

+2,414 ★ this week｜70,004 total｜Shell

The official Claude Code repo crossed 70,000 stars around its one-year anniversary (created 2025-02-22). The week's main talking point wasn't a new feature—it was a 39-point HN thread about Claude Code's GitHub automatically closing issues after 60 days. Community reactions were mixed: some called it reasonable issue triage; others argued it makes bug tracking unreliable.

6,740 open issues at time of writing reflects both the tool's market scale and the depth of real-world usage.

📈 #6 — google-research/timesfm｜Research Model Becomes an Office Tool via Google Sheets

TimesFM (Time Series Foundation Model) — a pretrained time-series foundation model for zero-shot forecasting.

+1,903 ★ this week｜9,725 total｜Python｜Apache-2.0

TimesFM itself isn't new, but the spike has a clear cause: on February 16, Google announced TimesFM integration into Connected Sheets (Google Workspace), letting business users run time-series forecasts directly inside Google Sheets—no SQL, no Python, no model training required.

That integration opened a research model that previously required ML expertise to financial analysts, supply chain planners, and business analysts overnight. A textbook example of research-to-product commercialization.

📈 #7 — Stremio/stremio-web｜A 2018 Streaming Client Unexpectedly Goes Viral

Stremio - Freedom to Stream

+1,087 ★ this week｜10,104 total｜JavaScript｜GPL-2.0

The hardest entry on this week's chart to explain. Stremio is an open-source media streaming client created in 2018. It jumped +1,087 stars this week with no identifiable driving event in GitHub or HN data. Possible causes: concentrated discussion in a community (Reddit? A Telegram channel?) or a feature update attracting users from the torrent ecosystem (Stremio supports external add-ons including Torrent). An open question—if you know the reason, let us know.

📈 #8 — muratcankoylan/Agent-Skills-for-Context-Engineering｜A Skill Library for Context Engineering

A comprehensive collection of Agent Skills for context engineering, multi-agent architectures, and production agent systems.

+1,072 ★ this week｜10,418 total｜Python｜MIT

Alongside obra/superpowers and huggingface/skills, this forms the week's Skills triangle. The focus is "Context Engineering"—how to design and manage AI agent context, covering multi-agent delegation, context compression strategies for production environments, and debugging methodologies.

If you're building complex AI agent systems and context management is your bottleneck, this is the week's most technically relevant repository to read.

📈 #9 — cloudflare/agents｜Stateful AI Agents on the Edge via Workers

Build and deploy AI Agents on Cloudflare

+940 ★ this week｜4,215 total｜TypeScript｜MIT

Cloudflare's official AI Agent framework for building and deploying stateful agents on the Workers platform, using Durable Objects for state persistence. With cloudflare/vinext (see New Repos below) also charting this week, the combined picture is clear: Cloudflare is assembling a complete edge AI application stack—agents for logic, vinext for the Next.js-compatible UI layer.

📈 #10 — SynkraAI/aios-core｜An OS-Layer Framework for AI-Driven Full Stack Development

Synkra AIOS: AI-Orchestrated System for Full Stack Development - Core Framework v4.0

+707 ★ this week｜1,805 total｜JavaScript

A full-stack development framework that positions AI agents as the central orchestrator, claiming 40–70% reduction in LLM token waste. The GitHub homepage links to allfluence/aios-core; the HN data doesn't surface verifiable third-party validation. On the chart this week, but lacks independently verifiable benchmarks. Run your own tests before relying on the token-savings claims.

Spotlight — Top New Repos Top 10

🆕 #1 — cloudflare/vinext｜AI-Written Next.js Alternative, $1,100 Development Cost in One Week

Vite plugin that reimplements the Next.js API surface — deploy anywhere

2,172 total ★｜TypeScript｜MIT｜Created 2026-02-24

Background: Next.js build output is tightly coupled to Vercel's infrastructure. OpenNext, the community alternative, adapts the output of next build—but is fragile because any Next.js update to internal APIs can break it.

vinext takes a different approach: it reimplements the stable public API of Next.js (App Router, Pages Router, middleware, server actions, streaming, ISR) on top of Vite, bypassing Vercel's internals entirely. Technical claims: 94% API coverage, 4.4× faster builds, 57% smaller bundles.

The most striking detail: a Cloudflare engineer directed Claude AI through 800+ coding sessions over 7 days, spending approximately $1,100 in API costs to write nearly the entire codebase. The Cloudflare blog post covers this in full—the project itself is a real-world AI coding case study.

Still experimental. HN discussion questioned whether the Next.js API surface is worth reimplementing at all. The U.S. government's CIO.gov site is already running it in production.

🆕 #2 — Leonxlnx/taste-skill｜Stop Your AI From Generating Generic-Looking UIs

Taste-Skill (High-Agency Frontend) — gives your AI good taste. Stops the AI from generating boring, generic, "slop"

1,524 total ★｜Skills framework｜Created 2026-02-19

A single SKILL.md file. Install it in Claude Code and it instructs the AI to ban: the AI purple/blue color palette, cliché copy like "Elevate/Seamless/Unleash," generic brand names like "Acme/Nexus/SmartFlow," and pure black #000000—while enforcing high-contrast neutral bases (Zinc/Slate) for all frontend UI generation.

One-line pitch: an AI aesthetics correction tool against vibe-coding slop. 1,524 stars in a week shows how many AI-assisted frontend developers share the same frustration.

🆕 #3 — ShinMegamiBoson/OpenPlanter｜Open-Source Palantir for Civic Oversight

1,310 total ★｜Python｜MIT｜Created 2026-02-20

OpenPlanter is a recursive LLM investigation agent with a terminal UI. It ingests corporate registries, campaign finance records, lobbying disclosures, and government contracts; resolves entities across datasets; and surfaces non-obvious connections through evidence-backed analysis. Default max recursion depth: 4 levels, with parallel subagent execution.

The author @shinboson frames it as: "so you can keep tabs on your government since they're almost certainly keeping tabs on you." MarkTechPost has detailed coverage.

🆕 #4 — anthropics/financial-services-plugins｜Anthropic's Official Finance Plugins

905 total ★｜Python｜Apache-2.0｜Created 2026-02-23

Ten official, open-source plugins released by Anthropic on February 24, built for Claude Cowork (Anthropic's enterprise agent platform, distinct from Claude Code). Coverage spans investment banking, equity research, private equity, and wealth management: DCF models, LBO models, comp analysis, CIM drafts, earnings updates, initiating coverage reports.

Integrated data providers include Daloopa, Morningstar, S&P Global, FactSet, PitchBook, Bloomberg, and others. Plugins are markdown files—fully forkable and customizable. A companion knowledge-work-plugins repo covers general knowledge workers (HR, design, etc.).

Bloomberg coverage here.

🆕 #5 — RightNow-AI/picolm｜A 1-Billion-Parameter LLM on a $10 Board

Run a 1-billion parameter LLM on a $10 board with 256MB RAM

882 total ★｜C｜MIT｜Created 2026-02-19

The core of this repo is ~2,500 lines of C11, zero dependencies, single binary around 80KB. Primary target hardware: Sipeed LicheeRV Nano ($10 RISC-V board, 256MB RAM) and the Raspberry Pi series.

Key technical specs: runtime RAM usage ~45MB (including ~40MB FP16 KV cache); model disk footprint 638MB (memory-mapped, streamed one layer at a time to fit in constrained RAM); supports TinyLlama 1.1B and any LLaMA-architecture model in GGUF format. Approximately 8–10 tokens/sec on Pi 4.

Pair it with openclaw/picoclaw (a Go orchestrator that pipes prompts via stdin/stdout to picolm as a subprocess) and you get a fully offline AI agent—no cloud, no API keys, no monthly subscription. Ideal for privacy-sensitive workloads or edge deployments without network access.

🆕 #6 — olvvier/apple-silicon-accelerometer｜Your MacBook Has a Hidden Accelerometer Nobody Told You About

reading the undocumented mems accelerometer + gyroscope on apple silicon macbooks via iokit hid

797 total ★｜Python｜MIT｜Created 2026-02-19

This repo reveals something that excited the hardware community: every Apple Silicon MacBook (M1 through M5) contains an undocumented MEMS accelerometer and gyroscope, accessible via IOKit HID at AppleSPUHIDDevice (vendor usage page 0xFF00), sampling at up to 800Hz. Apple provides no public API for it.

Highlights from the HN 152-point discussion: commenters connected it to Apple's older Sudden Motion Sensor (2005–2012), which protected spinning hard drive heads from drops; the current hardware likely serves Apple's "Vehicle Motion Cues" accessibility feature (mitigating motion sickness in moving vehicles). Someone experimented with resting their wrists on the trackpad and detected their own heartbeat via ballistocardiography—mechanical vibrations from cardiac output transmitted through the arms into the chassis. The overall tone was curiosity rather than privacy alarm.

🆕 #7 — Polymarket/polymarket-cli｜Official Polymarket CLI (Rust)

770 total ★｜Rust｜Created 2026-02-24

The official command-line tool from Polymarket, written in Rust. No other description provided. Among this week's flood of Kalshi/Polymarket tools, this is the only one from an official account—signaling that Polymarket is actively investing in its own CLI ecosystem.

🆕 #8 — Panniantong/Agent-Reach｜Function Unconfirmed

731 total ★｜Python｜MIT｜Created 2026-02-24

No official description. HN matches returned unrelated results. Unable to confirm the intended use case—check the repo directly before drawing conclusions.

🆕 #9–10 — Kalshi-Claw, Kalshi-Polymarket-Ai-bot｜Security Warning

⚠️ Security Warning: Both repos show highly abnormal stars-to-forks ratios (Kalshi-Claw: 690 ★ with only 8 forks; Kalshi-Polymarket-Ai-bot: 680 ★ with only 4 forks), strongly suggesting star inflation. Similar repos in this cluster have been documented by Permiso Security as containing malicious code (remote code execution, credential theft). If you're evaluating any Kalshi or Polymarket AI trading repo, conduct a full code review before running anything. Do not execute unknown trading agents against live accounts.

Monthly Trend Comparison

This week's only monthly holdover: obra/superpowers (🔁)

obra/superpowers has been on the monthly trending chart since mid-January. That means it's not riding a single media wave or viral tweet—it's in sustained word-of-mouth growth with genuinely new users discovering it every week. Against the backdrop of Jesse Vincent's background (Perl, Keyboardio), this looks less like hype and more like a practitioner with deep engineering instincts systematizing an AI coding methodology that actually works.

Weekly Trend Insights

Skills ecosystem shifts from personal tools to platform standard

The simultaneous appearance of four repos (obra/superpowers, huggingface/skills, muratcankoylan/Agent-Skills-for-Context-Engineering, Leonxlnx/taste-skill) isn't coincidence. It marks an inflection point: "prompt engineering" in AI coding is becoming "skill engineering," with platforms (HuggingFace) now providing official registries and individual developers packaging domain-specific skills (frontend aesthetics, context management). The question to watch: who builds the npm for Skills?

Embedded AI infrastructure taking quiet shape

alibaba/zvec (vector DB embedded in your process) and RightNow-AI/picolm (LLM on a $10 board) point in opposite directions technically but share the same core signal: AI infrastructure is moving from "cloud service" toward "embedded application." The SQLite analogy for zvec is right—like SQLite, its real competitive edge is zero ops, zero latency, zero cost, not benchmark numbers. This trend matters most to developers building privacy-sensitive or offline applications.

OpenClaw offspring: ecosystem creativity meets new security risks

OpenClaw (formerly Clawdbot, 100k stars in one week, renamed after Anthropic's trademark complaint) left a heavy footprint in this week's New Repos—especially prediction market tools. This is a double-edged story: the Skills framework clearly unleashes community creativity, but Permiso's security research has documented malicious repos mixed into the ecosystem, including credential theft and remote code execution. Do a full code review before running any unknown AI trading bot repo.