GitHub Hot — 26 March 2026

Top 10 repos trending on GitHub this week — what they do, why they matter, and how to use them in your projects.

1. slavingia/skills

3,592 stars this week · various

A Claude Code plugin that installs 10 business-strategy slash commands based on Sahil Lavingia's Minimalist Entrepreneur framework, turning book advice into interactive AI workflows.

Use case

Developers who want to productize a side project often stall at non-technical decisions — pricing, finding first customers, scoping an MVP. This repo wires those decisions into Claude Code as guided commands, so instead of vague AI chat you run /pricing and get a structured pricing conversation anchored to Lavingia's specific methodology. For example, Henry could run /mvp before building any new blog feature to force himself to scope it to the smallest shippable thing.

Why it's trending

Claude Code's plugin/skills marketplace just became a real ecosystem this week, and Sahil Lavingia (Gumroad founder) publishing his own skills repo is a high-signal proof-of-concept that's drawing attention to what's now possible — it's trending because it's one of the first 'thought leader publishes a plugin' moments for the platform.

How to use it

Open Claude Code in your project terminal (requires Claude Code CLI installed and authenticated).,2. Run /plugin marketplace add slavingia/skills then /plugin install minimalist-entrepreneur — Claude Code fetches and registers all 10 commands automatically.,3. Use contextual commands as decision gates: before writing any new feature, run /mvp and paste your idea; before launching, run /validate-idea with your problem statement.,4. Use /minimalist-review as a final sanity check on any product decision — paste your plan and get a framework-aligned critique.,5. Chain commands sequentially: /find-community → /validate-idea → /mvp mirrors the book's actual chapter progression, giving you a structured build journey inside your dev environment.

How I could use this

Run /validate-idea against every new blog post series or content category Henry considers — treat it as a forcing function to confirm there's a real audience before investing writing time, then publish the validation output as a 'why I'm writing about X' meta-post to demonstrate product thinking.
Use /pricing to build a transparent 'how I priced my services' case study page on the blog — paste his freelance/consulting rate card into the command, document the AI's structured pushback, and publish the before/after as a portfolio piece that signals business maturity to hiring managers.
Fork the repo and create a custom skills plugin for his blog's own Claude Code environment — add a /new-post skill that enforces his personal writing checklist (SEO slug, meta description, internal links, Supabase draft entry) so every post creation runs through a consistent AI-guided workflow he can demo as a technical blog post.

2. zarazhangrui/codebase-to-course

1,729 stars this week · various

A Claude Code skill that reverse-engineers any codebase into a self-contained, interactive HTML course — scroll nav, quizzes, animated diagrams, and plain-English code translations included.

Use case

When you inherit or fork a complex codebase (say, a Next.js SaaS starter or an open-source AI agent framework) and need to actually understand its architecture before modifying it, this generates a structured course from the source itself. Instead of reading through 40 files manually, you point it at a repo and get an interactive walkthrough that explains data flow, which files change for which features, and why key decisions were made — in plain English alongside the actual code.

Why it's trending

It hit virality this week riding the 'vibe coding' wave — it directly serves the massive cohort of non-CS developers building real products with AI tools who now face the 'I shipped it but I can't maintain it' wall. It also showcases Claude's extended thinking/tool use in a highly visual, shareable output format, which makes it a great demo artifact.

How to use it

Install Claude Code and make sure you have API access (claude.ai/code or the CLI). 2. Clone the repo: git clone https://github.com/zarazhangrui/codebase-to-course && cd codebase-to-course. 3. Copy the skill prompt (found in the README or the .claude skill file) into a new Claude Code session. 4. Run it against your target repo by providing the path: tell Claude something like 'Use the codebase-to-course skill on /path/to/my-nextjs-blog'. 5. Claude generates a single course.html file — open it in any browser, no server needed, and navigate with scroll or arrow keys.

How I could use this

Generate a public 'How This Blog Works' course from your own blog's repo and link it in your About page — it's a unique portfolio differentiator that shows both your code quality and your ability to communicate architecture to non-engineers.
Run this against popular AI/LLM libraries (like Vercel AI SDK or LangChain.js) that Henry is using in his blog's AI features, then screenshot the animated component diagrams and embed them in technical blog posts as visual aids — instantly making posts more shareable.
Build a 'Codebase Explainer' micro-tool as a blog feature: Henry pastes a GitHub URL, a serverless function clones the repo and runs the Claude skill via the API, and the resulting HTML course is stored in Supabase Storage and served back — letting readers generate courses for any public repo directly from his site.

3. dontbesilent2025/dbskill

1,642 stars this week · various

A Claude Code skill pack distilled from 12,307 tweets into structured business diagnosis workflows — installable as slash commands that chain together like a decision tree.

Use case

When you're using Claude Code for more than coding — writing, strategy, content — you have no reusable prompt infrastructure. Every session you re-explain context from scratch. This repo solves that by giving you installable slash commands (e.g., /dbs-content, /dbs-hook) that carry methodology baked in, so /dbs-diagnosis can automatically hand off to /dbs-action when it detects an execution problem rather than a structural one.

Why it's trending

Claude Code's custom skill/command system just became a serious workflow layer for knowledge workers, and this is one of the first public examples of someone packaging domain expertise (business diagnosis, content strategy) as a distributable skill set — it's a template for what 'prompt engineering as a product' looks like in practice.

How to use it

Install the skill pack into your Claude Code environment: npx skills add dontbesilent2025/dbskill — or manually clone and copy into ~/.claude/skills/,2. Open Claude Code in any project directory and type /dbs — the router will ask what problem you're working on and route to the right sub-skill automatically,3. For blog content specifically, run /dbs-content with a draft post pasted in — it runs a 5-dimension diagnostic and flags structural issues like weak hooks or unclear value propositions,4. If it flags a hook problem, it will recommend /dbs-hook — run that to generate 10+ alternative opening lines for the post based on short-video hook methodology,5. Study the 知识库/原子库/atoms.jsonl file — 4,176 structured knowledge atoms with confidence scores and topic tags — to understand how to build your own skill knowledge base from existing content

How I could use this

Run /dbs-content as a pre-publish gate for every blog post: pipe your draft markdown into Claude Code before pushing to Supabase, and only publish if it passes the 5-dimension content diagnostic — log the diagnostic results to a post_quality_scores table to track your writing improvement over time
Build a custom /henry-career skill modeled on the dbskill pattern: extract methodology from your own GitHub activity, blog posts, and project READMEs into a personal atoms.jsonl, then create slash commands like /resume-fit [job-description] that use your distilled 'career atoms' as context for tailored cover letter generation
Clone the chatroom pattern (Hayek × Mises × Claude multi-voice dialogue) to build a /debate-post skill for your blog: given a draft opinion post, spin up two AI personas representing opposing viewpoints and have them interrogate your argument — surface the counterarguments you missed before a real reader does, and optionally publish the debate transcript as a companion piece

4. HKUDS/OpenSpace

1,261 stars this week · Python

OpenSpace is a shared memory and skill-evolution layer that lets AI agents (Claude Code, Codex, Cursor, etc.) learn from past task solutions and share that knowledge across sessions and agents — cutting token usage by ~46%.

Use case

Every time you run an AI coding agent on a new task, it reasons from scratch — burning tokens rediscovering patterns it already solved last week. OpenSpace intercepts agent workflows, stores successful task-solution pairs in a shared vector store, and retrieves relevant prior solutions as context before the agent starts reasoning. Concrete example: your Claude Code agent debugs a Supabase RLS policy issue, OpenSpace logs the fix; next time Cursor hits the same pattern, it gets the solution injected automatically instead of spending 20 API calls re-discovering it.

Why it's trending

It dropped this week riding the wave of Claude Code and Codex CLI going mainstream — developers are now running agentic coding loops in CI and feeling the token bill, so a drop-in optimization layer that promises 46% savings with zero model changes is immediately actionable. The '$11K earned in 6 hours' hook from a hackathon win also drove significant social traction.

How to use it

Install: pip install openspace-agent (Python 3.12+) and set your LLM API keys in env.
Run the CLI against your existing agent: openspace --query "refactor the auth module to use Supabase SSR cookies" — OpenSpace wraps the agent call, queries its skill store first, and injects relevant past solutions as context.
On task completion, OpenSpace auto-logs the successful solution pattern to its local/shared vector store (configurable: local SQLite or cloud at open-space.cloud).
Enable cross-agent sharing by pointing multiple agents (e.g., Claude Code in terminal + Cursor in IDE) at the same OpenSpace instance via OPENSPACE_STORE_URL env var — they now share a collective memory.
Inspect evolved skills: openspace skills list to audit what patterns have been learned and prune noise.

How I could use this

Wire OpenSpace into your blog's content-generation pipeline: every time your AI agent drafts a post (summarizing a paper, generating code snippets), log the successful prompt-output pairs. After 20 posts, the agent retrieves your established voice and structure patterns automatically — zero prompt engineering per session.
Build a career-tool agent (resume tailoring, cover letter generation) that uses OpenSpace to accumulate a personal skill store of winning phrasings per job category. After Henry applies to 10 jobs, the agent stops re-reasoning about 'how to frame a Next.js project for a fintech role' and retrieves the proven template instead, cutting generation cost and improving consistency.
Integrate OpenSpace as a middleware layer between your Supabase backend and an AI debug agent: when the agent resolves a Row Level Security bug, a failed Edge Function, or a schema migration conflict, the fix gets stored. Surface these as a living 'lessons learned' page on the blog — auto-generated from the OpenSpace skill store via a nightly cron that pulls new entries and publishes them as short posts.

5. louislva/claude-peers-mcp

1,255 stars this week · TypeScript

An MCP server that lets multiple Claude Code instances running in parallel terminals discover each other and exchange messages in real time, enabling multi-agent coordination on a single machine.

Use case

When you're running Claude Code in 3 separate terminals — one refactoring your Supabase schema, one writing API routes, one updating frontend components — they have no awareness of each other and can stomp on shared files or make conflicting assumptions. claude-peers gives each instance a discovery mechanism and message bus so Claude A can ask Claude B 'are you touching auth.ts right now?' before making changes, effectively turning isolated agents into a loosely coordinated team.

Why it's trending

Multi-agent Claude Code workflows went mainstream this week after Anthropic shipped Claude Code with expanded tool-use and MCP support, and developers are immediately hitting the coordination problem when running parallel sessions on large codebases. This repo is the first practical solution to that specific pain point with zero infrastructure.

How to use it

Clone and install: git clone https://github.com/louislva/claude-peers-mcp.git ~/claude-peers-mcp && cd ~/claude-peers-mcp && bun install
Register globally: claude mcp add --scope user --transport stdio claude-peers -- bun ~/claude-peers-mcp/server.ts
Add a shell alias to avoid retyping flags: alias claudepeers='claude --dangerously-load-development-channels server:claude-peers'
Open two terminals and start claudepeers in each — the broker daemon auto-starts on first launch.
In terminal 1, prompt Claude: 'List all peers on this machine' — you'll see terminal 2's session with its working directory and git context, then send a message: 'Send a message to peer [id]: what files are you editing?'

How I could use this

Run one Claude peer dedicated to blog content generation (MDX drafting, slug creation, tag suggestions) and a second peer handling Supabase schema/API work — have the content peer message the API peer to auto-generate a matching REST endpoint or RLS policy whenever a new post type is created, keeping data model and content in sync without manual coordination.
Split your portfolio's AI resume-matching pipeline across two peers: peer A ingests a job description and scores your experience against it, peer B simultaneously rewrites the relevant bullet points — have peer A message peer B with the low-scoring sections so the rewrite is targeted, then a third peer assembles the final output. This parallelizes what is currently a sequential prompt chain.
Use two peers during blog feature development as a lightweight code review loop: peer A writes the feature (e.g., a new AI-generated post summary component), then messages peer B with the file path and a prompt like 'review this for accessibility and TypeScript strictness' — peer B responds inline, and peer A applies the fixes before you ever open a PR. Effectively a free async pre-review agent running locally.

6. dou-jiang/codex-console

1,039 stars this week · Python

A Chinese-community-maintained Python console for automating OpenAI account registration, token harvesting, and bulk account management — essentially a patched fork that keeps working as OpenAI tightens its signup flow.

Use case

If you're running an AI proxy service or LLM API aggregator (like a self-hosted one-api/new-api instance) and need to bulk-provision OpenAI accounts to pool API access, this handles the brittle parts: Sentinel POW solving, OAuth token refresh, email verification polling, and subscription status checks. Concrete example: you run a small team that shares a pooled OpenAI quota via new-api, and you need to onboard 20 accounts without manually clicking through each signup flow.

Why it's trending

OpenAI recently hardened its registration flow with Sentinel POW challenges and changed how post-signup token issuance works, breaking every existing automation script — this repo is one of the first publicly patched forks that actually handles the new flow, making it immediately useful to the large Chinese developer community running LLM proxy infrastructure.

How to use it

Clone and install deps: git clone https://github.com/dou-jiang/codex-console && cd codex-console && pip install -r requirements.txt (requires Python 3.10+). 2. Configure your email provider (Outlook or CloudMail) and proxy settings in the config file — the proxy_url field in the CPA config is mandatory if you're outside China or need IP rotation. 3. Launch the web UI: python main.py then open http://localhost:8000 to access the dashboard. 4. Use the bulk registration flow to queue accounts, monitor logs in the built-in log viewer, and export results in the Codex account format for downstream import. 5. Point the newApi upload target at your one-api/new-api instance endpoint so harvested tokens auto-populate your LLM proxy pool.

How I could use this

Build a personal API cost dashboard on Henry's blog: use the exported account/quota data from codex-console as a data source, pipe it into a Supabase table via a cron job, and render a live 'API budget remaining' widget on the blog's admin panel — useful for transparently showing readers what it costs to run AI features.
Create a 'token health monitor' career tool: if Henry is building AI-powered resume or cover letter generators, wrap the OAuth token refresh logic from this repo into a lightweight Python microservice that auto-rotates OpenAI tokens before they expire, so his career tools never fail mid-request for paying users.
Use the Sentinel POW solver as a reference implementation: the POW-solving logic in this repo is one of the few public Python implementations of OpenAI's current challenge — study it to build a robust retry/fallback layer in Henry's Next.js API routes that gracefully handles OpenAI auth errors instead of returning 500s to blog readers using his AI features.

7. alvinunreal/awesome-opensource-ai

960 stars this week · various · agents ai artificial-intelligence awesome

A hand-curated index of production-ready open-source AI tools across the full stack — models, RAG, agents, MLOps, and infra — filtered for projects that are actually open-source (not just open-weight).

Use case

When building an AI feature, you waste hours evaluating whether to use LangChain vs LlamaIndex vs Haystack for RAG, or which vector DB won't lock you in. This repo cuts that research time by grouping battle-tested OSS alternatives by category with GitHub star counts as a quality signal. For example: Henry needs a self-hosted embedding + vector search stack — he can scan section 5 (RAG & Knowledge) and section 3 (Inference) in 10 minutes instead of Googling for an hour.

Why it's trending

The wave of 'open-weight but closed-license' models (Llama 2 commercial restrictions, etc.) has developers actively seeking truly open alternatives, making a license-aware curated list acutely valuable right now. It's also hitting 1k stars fast because it covers the full MLOps stack, not just models — filling a gap that awesome-llm lists miss.

How to use it

Clone or bookmark the repo: git clone https://github.com/alvinunreal/awesome-opensource-ai — then open the README and use Ctrl+F to jump to the category matching your current build task (e.g. 'RAG', 'Agents', 'Inference').,2. For each candidate tool, click through to its GitHub page and check: stars trajectory (use star-history.com), last commit date, and license (Apache 2.0 / MIT = safe for commercial use).,3. Cross-reference section 8 (MLOps/LLMOps) when you're ready to move a prototype to production — tools like Langfuse, Phoenix, or Helicone for tracing are listed there and integrate directly with LangChain/LlamaIndex.,4. Use the Contributing guide to submit tools you discover — good for visibility in the OSS community and building a public track record.,5. Pin the sections most relevant to your stack (e.g. section 12 for self-hosted UIs, section 5 for RAG) as browser bookmarks for fast reference during architecture decisions.

How I could use this

Use section 5 (RAG & Knowledge) to pick a self-hosted stack (e.g. Qdrant + Ollama + LlamaIndex) for a blog-search feature that lets readers semantically search all of Henry's posts — avoiding OpenAI API costs and data privacy concerns entirely.
Cross-reference section 8 (MLOps) to add LLM observability (e.g. Langfuse, which is listed) to any AI career tool Henry builds — logging prompt/response pairs to Supabase for fine-tuning data collection while also debugging hallucinations in a resume matcher.
Use section 4 (Agentic AI) to evaluate open-source agent frameworks (e.g. AutoGen, CrewAI) for building a 'blog post drafting agent' that takes a topic, auto-researches via web search tool, and outputs a structured draft — all self-hosted so Henry owns the workflow and can write about the build publicly.

8. GAIR-NLP/daVinci-MagiHuman

930 stars this week · Python

daVinci-MagiHuman is a fully open-source 15B single-stream transformer that generates synchronized talking-head videos with audio from text in seconds, rivaling closed commercial models.

Use case

The real problem: generating realistic, lip-synced human presenter videos without paying per-minute API fees to HeyGen or Synthesia. Concrete scenario: you have a blog post and want to auto-generate a 30-second 'author intro' video of a digital avatar reading the summary aloud — this model does that locally at 1080p in under 5 minutes on one H100, with accurate mouth/expression sync across 6 languages.

Why it's trending

It dropped this week as a fully open-source alternative to locked-down commercial video avatars (HeyGen, D-ID), and its benchmark win rates against LTX 2.3 and Ovi 1.1 are making the research community take notice — especially since the full model stack (base, distilled, SR) is on HuggingFace right now.

How to use it

Clone and install: git clone https://github.com/GAIR-NLP/daVinci-MagiHuman && pip install -r requirements.txt (requires Python 3.12+, PyTorch 2.9+, and an H100/A100 for 1080p).
Download the distilled model from HuggingFace: huggingface-cli download GAIR/daVinci-MagiHuman --local-dir ./checkpoints.
Run inference with a text prompt and reference portrait image:

from magihuman import MagiHumanPipeline
pipe = MagiHumanPipeline.from_pretrained('./checkpoints')
video = pipe(
    text='Hello, welcome to my blog post on RAG pipelines.',
    portrait='./henry_photo.jpg',
    duration=5,
    resolution='256p'  # use '1080p' for production
)
video.save('intro.mp4')

For 256p tests on a consumer GPU (RTX 4090), use the distilled model variant — it generates 5s clips in ~2s on H100, so expect ~15-20s on a 4090.
Integrate the output MP4 into your Next.js blog by uploading to Supabase Storage and serving via a <video> tag or embedding in a post header.

How I could use this

Auto-generate a 'Henry introduces this post' talking-head video for each blog article: pipe the post's AI-written summary + a headshot into MagiHuman, upload the MP4 to Supabase Storage, and embed it as a hero element — turns a static blog into a video-first experience without recording anything manually.
Build a 'AI mock interviewer' tool for your portfolio: use MagiHuman to pre-render a set of realistic interviewer avatar videos asking common technical questions, then serve them in a Next.js app where users practice answering — differentiates your portfolio from every other developer's static site.
Create a multilingual blog post reader feature: when a visitor selects Japanese or French from a language picker, trigger a Supabase Edge Function that calls a locally-hosted MagiHuman instance to generate an on-demand translated talking-head video of the post summary — demonstrating real full-stack AI integration in a live project.

9. wong2/weixin-agent-sdk

918 stars this week · TypeScript

A TypeScript SDK that bridges any AI agent (OpenAI, Claude, Codex, etc.) to WeChat messaging via a dead-simple Agent interface, making WeChat your chat UI for arbitrary AI backends.

Use case

Developers with existing AI backends have no easy way to expose them through WeChat without reverse-engineering Tencent's proprietary APIs. This SDK abstracts the WeChat connection layer so you implement one chat(request): Promise<ChatResponse> method and your AI is instantly available to any of WeChat's 1.3B users — e.g., wire your RAG-powered blog assistant into WeChat so readers can ask questions about your posts directly in a chat they already use daily.

Why it's trending

WeChat just opened 'Clawbot' (AI agent slots) to third-party developers, and this is the first clean TypeScript SDK to bridge that opening to the ACP (Agent Client Protocol) ecosystem — timing it perfectly with the Claude Code and Codex ACP adapters landing this week.

How to use it

Install the SDK: npm install weixin-agent-sdk
Implement the Agent interface with your AI logic:

import { login, start, type Agent } from 'weixin-agent-sdk';
import OpenAI from 'openai';

const client = new OpenAI();
const history = new Map<string, OpenAI.ChatCompletionMessageParam[]>();

const blogAgent: Agent = {
  async chat({ conversationId, text }) {
    const msgs = history.get(conversationId) ?? [];
    msgs.push({ role: 'user', content: text });
    const res = await client.chat.completions.create({
      model: 'gpt-4o',
      messages: [{ role: 'system', content: 'You are an assistant for Henry\'s blog.' }, ...msgs],
    });
    const reply = res.choices[0].message.content ?? '';
    msgs.push({ role: 'assistant', content: reply });
    history.set(conversationId, msgs);
    return { text: reply };
  },
};

await login(); // QR code scan in terminal
await start(blogAgent);

Run ts-node index.ts, scan the QR code with your WeChat account, and your agent is live.
For Claude Code specifically, skip all the above and just run npx weixin-acp claude-code — zero additional code needed.

How I could use this

Build a WeChat bot backed by a Supabase pgvector index of all your blog posts — readers message the bot a question and get back a cited answer with a link to the relevant post, driving traffic back to your Next.js blog without any app install friction.
Create a private WeChat career assistant that accepts a pasted job description, queries your Supabase-stored resume data, and replies with a tailored gap analysis and talking points — useful for your own job search and demonstrable as a live portfolio project in interviews.
Expose your blog's draft review workflow over WeChat: send a markdown draft to the bot, it calls GPT-4o to return SEO suggestions, readability score, and a proposed meta description directly in the chat thread — letting you do editorial QA from your phone without opening a browser.

10. opa334/darksword-kexploit

887 stars this week · Objective-C

A kernel-level exploit for iOS 15–26.0.1 reimplemented in Objective-C, enabling arbitrary kernel memory access on affected devices — relevant to jailbreak and security research communities.

Use case

This solves the problem of needing a reproducible, readable Objective-C implementation of the DarkSword kernel exploit for security research, jailbreak tool development, or CVE analysis. For example, a security researcher auditing iOS kernel mitigations could use this as a reference implementation to study how the vulnerability works before Apple patches it, or a jailbreak developer could build it as a stepping stone toward a full userland-to-kernel privilege escalation chain.

Why it's trending

iOS 26 (the newly branded name for what was expected to be iOS 19) was just announced at WWDC 2025, and a working kernel exploit covering up to 26.0.1 is an extremely high-value find that immediately draws attention from the jailbreak and mobile security research communities. The reimplementation in Objective-C (rather than C or a PoC script) makes it significantly more accessible and integrable into real tooling.

How to use it

Clone the repo: git clone https://github.com/opa334/darksword-kexploit and open the Xcode project — you'll need a Mac with Xcode and an Apple Developer account to sign and deploy to a device.
Note that offsets are currently hardcoded for iOS 15.x — before running on any other version, you must locate the correct kernel offsets for your target iOS version using a tool like iometa or joker on the IPSW.
Build and sideload to a non-production test device (never run kernel exploits on a primary device) using xcodebuild or Xcode's device runner.
Inspect the core exploit entry point in the Objective-C source to understand the vulnerability trigger — trace the kernel read/write primitives it establishes before attempting to extend it.
Cross-reference with the original DarkSword disclosure or any associated CVE to understand the attack surface (likely a type confusion or UAF in a kernel extension) before modifying offsets for newer iOS versions.

How I could use this

Write a deep-dive blog post titled 'Reading a Kernel Exploit: What Non-Security Devs Can Learn from DarkSword' — walk through the Objective-C source, explain what kernel read/write primitives are, and use it to teach memory safety concepts relevant to any systems programmer. This kind of technical breakdown drives serious SEO traffic from the security and iOS dev community.
Not directly applicable to resume/career tooling — this is a low-level iOS security repo with no overlap with Henry's Supabase/Next.js stack. Attempting to force a connection here would undermine credibility. Skip this category for this repo.
Use the public attention around this exploit as a hook for an AI-powered 'CVE Explainer' feature on your blog — pipe CVE descriptions and linked GitHub repos into an LLM (via OpenAI or Claude API) and auto-generate plain-English breakdowns. This exploit trending is a perfect real-world test case to demo the feature and publish as a blog post driving traffic while the topic is hot.