GITHUB HOT REPOS

Hot GitHub repos
of the week.

Report by The Next New Thing.

Scroll for the report ↓

b1 · BONUS · Anil-matcha/open-genjutsu

Put your own character into any short video clip

ELI5

Give it a short video clip and a few pictures of a character, and it remakes the clip with that character doing the same moves, then puts the original sound back on.

View on GitHub →
Stars
1.0k
Language
Python
Started
Oct 2026
GitHub repo card for Anil-matcha/open-genjutsu — click to open the repo
Highlights
  • Swap the performer

    Add a 4 to 15 second clip and up to eight character images; it renders a new take with your character copying the original moves.

  • Two render choices

    Genjutsu renders with Seedance 2.5; Genjutsu Lite uses Wan 3.0, which the project warns may not copy the performance frame for frame.

  • Original audio restored

    After the render, FFmpeg (a free video tool) puts the source soundtrack back and saves an MP4 on your computer.

Star growth over time for Anil-matcha/open-genjutsu
Anil Matcha (@matchaman11) launches Open-Genjutsu as a free alternative to paid video-cloning apps.

my experience

b2 · BONUS · JuneYaooo/rea-video

Copy videos by deconstructing shot by shot.

ELI5

Give it a video and your AI coding helper goes through it one shot at a time, telling you what was filmed, what was made with AI and what was built in code, plus a plan to remake it.

Stars
7.8k
Language
Python
Started
Oct 2026
GitHub repo card for JuneYaooo/rea-video — click to open the repo
Highlights
  • Filmed, AI or code, per shot

    It measures frame timing, image noise and hidden AI tags in the file to judge each shot, then flags conflicts for the agent to check by eye.

  • Captions, cuts and sound pulled out

    Every caption with its timing (exportable as subtitle files), each cut, the music tempo, and where each sound effect lands.

  • A remake plan you can follow

    It says what to film yourself, what to generate and with which AI video model, and what to rebuild as code with Remotion (a video-from-code tool).

Star growth over time for JuneYaooo/rea-video
Riley Brown (@rileybrown) says the REA-Video skill now copies an editor's style very well, and warns where that is heading.
b3 · BONUS · NYLLON-SOFTWARE/ember-cloudflare

Slack-style team chat that runs on your own Cloudflare

ELI5

Imagine a Slack-style chat for your team that lives in your own Cloudflare account (a web hosting service). One button sets it up, and a small team can run it for nothing.

View on GitHub →
Stars
1.0k
Language
TypeScript
Started
Oct 2026
GitHub repo card for NYLLON-SOFTWARE/ember-cloudflare — click to open the repo
Highlights
  • One-click deploy, $0 for a small team

    Runs on Cloudflare's free plan; the project measured about 100 people sending 30 messages a day. Past that, $5 a month for the whole account.

  • Keeps every message

    No 90-day cutoff like Slack's free plan. Messages and files stay in your own Cloudflare account for as long as you keep them.

  • Slack-ready bots plus an @ember assistant

    Tools that post to a Slack webhook can post to Ember instead. Mention @ember for answers from a model you choose, including free Workers AI ones.

Star growth over time for NYLLON-SOFTWARE/ember-cloudflare
The Kevin Rose story

Digg's co-founder built his own free Slack

Kevin Rose, who co-founded Digg and helped back Slack as a Google Ventures partner, spent a month building a team chat that anyone can run for free. He open-sourced it on Nov 9.

  • Who he is

    His own site lists Digg, Revision3, Zero, Oak and PROOF among his startups, and says that at Google Ventures he helped back Slack, Uber and Nextdoor.

  • What he posted

    On Nov 9, 2026 he wrote "Want your own Slack for $0? I open-sourced Ember." The post had 384,000 views, 2,587 likes and 3,280 bookmarks when we checked.

  • Why he built it

    His pitch is cost: it "deploys to @Cloudflare's free tier in about a minute." In a follow-up he added it is "Great for private communities," with Memberful, email-domain and invite sign-ups.

  • How it was built

    Asked on X (by @Jason) how long it took, he replied "4-5 hrs a day, for about a month. 968 commits :)". On GitHub, most commits are credited to the Claude coding agent's account.

Kevin Rose (@kevinrose) open-sources Ember: "Want your own Slack for $0?"
Kevin Rose's follow-up: built for private communities, with the GitHub repo linked.
sherlock-project · Username search across 400+ sites · MIT

Type one username, find its accounts on 400+ sites

ELI5

Give it a username and it checks more than 400 websites, from Reddit to GitHub, for an account with that exact name. It runs from your computer's command line (a text window for typing commands).

View on GitHub →
Stars
110.9k
Language
Python
Started
Dec 2018
GitHub repo card for sherlock-project/sherlock — click to open the repo
sherlock-project / sherlock

Ravi ran this

sherlock
$ sherlock andrewwarner
[*] Checking username andrewwarner on:
 
[+] DEV Community: https://dev.to/andrewwarner
[+] Discord: https://discord.com
 
[*] Search completed with 69 results
 
Go deeper than a username. Explore public profiles and export your findings.
Try OSINTSearch: https://osintsearch.org
Highlights
  • One command, 400+ sites

    Run sherlock with one or several usernames and it checks every site on its list at once, then saves the hits to a text file.

  • Exports to CSV and Excel

    Flags save results as CSV or Excel files, and --browse opens every found profile in your browser.

  • Proxy support, easy installs

    You can route searches through a proxy, and install it with pipx or Docker (a packaged app that runs anywhere).

Star growth over time for sherlock-project/sherlock
Trending GitHub Repositories (@trending_repos) names Sherlock repo of the day on Nov 8: 2,372 stars in 24 hours.
Nav Toor (@heynavtoor) explains Sherlock and why your reused username is a fingerprint (April 2026).
Louis Gleeson (@aigleeson) on what Sherlock does and does not do: public clues only, no password breaking (July 2026).
Jackson Mittag installs Sherlock and runs a live, ethical demo in about six minutes.
Chapters — click to jump
whoamitang shows how easy username lookups make stalking, using Sherlock (first half) and FinalRecon.
Chapters — click to jump
G MAN : Security runs username searches with Sherlock, then covers reducing your own digital footprint.
Chapters — click to jump
GitHub Issue · sherlock-project/sherlock #3207
A crafted username can write files outside the output folder

On Nov 9, 2026 a user reported that Sherlock builds result filenames straight from the username, so a name like ../poc_escape writes its report outside the chosen folder (a path traversal bug, CWE-22). A fix, pull request #3211, was opened the same day and is still unmerged.

Open issue #3207 on GitHub →
anthropics · Anthropic's skill library, relicensed · Apache-2.0 per skill

Anthropic's Office skills for Claude are now open source

ELI5

Give Claude a shelf of add-on abilities from the company that makes it. Each folder teaches one task, like building a PowerPoint deck, filling in a PDF form, or writing a brand-new skill of your own.

View on GitHub →
Stars
194.0k
Language
Python
Started
Sep 2025
GitHub repo card for anthropics/skills — click to open the repo
Highlights
  • Office skills went Apache 2.0 on Nov 6

    Plug these skills into your own proprietary software, internal tools, or non-Claude AI agents without licensing fees or usage restrictions.

  • Word, PowerPoint, PDFs, and Excel

    These skills give your agent the Python scripts to generate properly formatted Word documents, PowerPoint decks, PDFs, and Excel spreadsheets.

  • Example uses

    Use the skills/pdf blueprint to extract specific form fields from a PDF, or skills/xlsx to parse and update spreadsheet formulas.

Star growth over time for anthropics/skills
Nick Babich installs Anthropic's skill-creator from this repo and builds a new Claude skill in five minutes.
Chapters — click to jump
Debbie O'Brien installs skill-creator from this repo, reviews a real skill with it, and runs evals.
Chapters — click to jump
milind-soni · Open-source Grok Bot alternative · Apache-2.0

A free "Grok Bot."

ELI5

Imagine a messaging app where each contact is an AI helper with its own job, its own computer and its own connected apps. You chat with them, watch them work, and approve the steps that matter.

View on GitHub →
Stars
22.0k
Language
TypeScript
Started
Aug 2026
GitHub repo card for milind-soni/OpenMausBot — click to open the repo
Highlights
  • Runs on the AI logins you already have

    Bots use the Claude Code, Codex and Grok command-line tools already signed in on your machine, so your existing subscriptions do the work.

  • Each bot gets its own computer

    A bot can use a cloud Linux desktop from Boat (paid after a trial), an isolated local virtual machine, or your own Mac after you opt in.

  • 500+ apps and event triggers

    Connect Gmail, Slack, Notion and more through Composio, then have a bot start work when an email or app event arrives.

Star growth over time for milind-soni/OpenMausBot
typecraft (@typecraft_dev), a fan of locally run tools, points people who want off Grok Bot to OpenMausBot.
Guilherme Vasques (@Ga_Vasques) compares MausBot with Codync, Rakazo and OpenBot as Grok Bot alternatives and ranks them.
Co-builder Omkar Satpute (@BuildwithOmkarr) introduces OpenMausBot, quote-posting CopilotKit's OpenMuse launch with near-identical wording.
Full Stack walks through install, each bot's computer, approval cards, memory, triggers, and what you give up versus xAI's Grok Bot.
Chapters — click to jump
Prism Labs explains what OpenMausBot does, how each bot gets a computer, and the free app versus paid cloud split.
Chapters — click to jump
Julian Goldie SEO demos OpenMausBot in August, two days after launch: picking a model per bot, computer use, and 500+ apps.
Chapters — click to jump
vercel-labs · A web browser for AI agents · Apache-2.0

A web browser your AI agent drives one command at a time

ELI5

Give your AI coding helper its own web browser. It reads each page as a short list of buttons and links, clicks or types, then checks what changed, so it can test your site or fill in forms for you.

View on GitHub →
Stars
50.5k
Language
Rust
Started
Jan 2026
GitHub repo card for vercel-labs/agent-browser — click to open the repo
Highlights
  • Saved flows, new on Nov 10

    Record a browser task once, then replay it. Version 0.42.0 tries a fast direct request first and falls back to replaying the clicks. Experimental.

  • Pages as short text, not screenshots

    A snapshot lists each button or link with a tag like @e1, and the agent acts on the tag. Its built-in guide now tells agents to prefer this over screenshots.

  • Plugs into many agents and browsers

    Run it from the terminal or as an MCP server (a plug-in port for AI tools), with Chrome, iPhone Safari in a simulator, or cloud browsers like Browserbase.

Star growth over time for vercel-labs/agent-browser
Mario Nawfal's Roundtable (@RoundtableSpace) explains how agent-browser lets a coding agent test the site it just built, in three install steps.
Hung Vinh (@hungv47) calls agent-browser 'arguably the best browser tool on the market', then posts his own test where OpenCode Browse beat it.
Ryan Carson (@ryancarson) pairs Codex with agent-browser and agentation to fix UI by clicking the element you want changed.
Better Stack demos agent-browser fixing form validation in a React app and compares it with Browser Use and Playwright MCP. Filmed in January; core commands still match.
Chapters — click to jump
Cole Medin switches from Playwright MCP to agent-browser and has it validate a full app live. Filmed in January; the snapshot-and-click loop is unchanged.
Chapters — click to jump
ZhuLinsen · AI stock briefings, run free · MIT

AI writes a daily report on your stock watchlist

ELI5

Give it a list of stocks and an AI model key, and every weekday it gathers prices and news, then sends you a short buy, watch, or sell summary on Telegram, Slack, or email.

View on GitHub →
Stars
71.3k
Language
Python
Started
Jan 2026
GitHub repo card for ZhuLinsen/daily_stock_analysis — click to open the repo
ZhuLinsen / daily_stock_analysis

Ravi ran this on NVDA

report.md
$ python main.py --stocks NVDA
# 🎯 2026-11-13 Decision Dashboard
 
> Analyzed **1** stocks | 🟢Buy:0 🟡Watch:0 🔴Sell:1
Market status: US · Intraday
 
## 📊 Summary
 
🟠 **NVIDIA Corporation(NVDA)**: Reduce | Score 35 | Bearish
 
---
 
## 🟠 NVIDIA Corporation (NVDA)
 
### 📰 Key Updates
 
**💭 Sentiment**: Neutral sentiment based on technicals.
**📊 Earnings Outlook**: N/A
 
**🚨 Risk Alerts**:
- Strong bearish trend, be cautious of further declines
- Volume indicates a pullback, but not a strong buy signal
 
**✨ Positive Catalysts**:
- Event-led, technicals not yet confirmed
 
**📢 Latest News**: No recent news found for NVIDIA Corporation.
 
### 📌 Core Conclusion
 
**🟠 Reduce** | Bearish
 
> **One-line Decision**: Reduce position due to bearish trend and strong bias.
 
⏰ **Time Sensitivity**: Act now
 
| Position | Action |
|---------|---------|
| 🆕 **No Position** | Do not initiate a position. |
| 💼 **Holding** | Reduce position size and consider taking profits. |
 
### 📈 Market Snapshot
 
| Close | Prev Close | Open | High | Low | Change % | Change | Amplitude | Volume | Turnover |
|------|------|------|------|------|-------|-------|------|--------|--------|
| 160.40 | 173.73 | 169.89 | 170.00 | 159.21 | -4.54% | -13.33 | 6.21% | 1.46 亿股 | N/A |
 
| Price | Volume Ratio | Turnover Rate | Source |
|-------|------|--------|----------|
| 160.40 | N/A | N/A | Fallback |
 
### 📊 Data View
 
**MA Alignment**: Bullish alignment not met; MA5 > MA10 > MA20 is not satisfied. | Bullish Alignment: ❌ No | Trend Strength: 35/100
 
| Price Metrics | Price |
|---------|------|
| Price | 160.39999389648438 |
| MA5 | 169.57599792480468 |
| MA10 | 182.5 |
| MA20 | 198.56600112915038 |
| Bias (MA5) | -5.41% Safe |
| Support | N/A |
| Resistance | 222.19000244140625 |
 
**Volume**: Volume Ratio N/A (Normal volume) | Turnover Rate N/A%%
💡 *Volume is normal, but it is a pullback, indicating selling pressure.*
 
**Chip Structure**: Chip distribution is disabled or temporarily unavailable; chip signals were not used.
 
### 🛡️ Phase Decision Guardrail
 
| Action Window | Current Action | Next Check |
|---------|---------|---------|
| Intraday tracking | Act now | Next checkpoint or market-local time |
 
**Watch Conditions**:
- Watch for further bearish signals or volume confirmation
 
**Confidence Reason**: Confidence level is medium due to bearish trend and strong bias, but no major negative news.
 
**Data Limitations**:
- Volume and chip data are partial or missing
- quote: fallback
- technical: partial
 
### 🎯 Battle Plan
 
**📍 Action Levels**
 
| Action Levels | Price |
|---------|------|
| 🎯 Ideal Entry | N/A |
| 🔵 Secondary Entry | N/A |
| 🛑 Stop Loss | 155.00 (break below MA20 or 5% below current price) |
| 🎊 Target | 165.00 (previous high/round-number level) |
 
**💰 Position Size**: 0/10
- Entry Plan: Avoid initiating new positions due to bearish trend.
- Risk Control: Set stop-loss at 155.00 to protect against further declines.
 
**✅ Checklist**
 
- ✅ Check 1: bullish alignment - Not met
- ⚠️ Check 2: reasonable bias - Bias is -5.41% (safe range)
- ✅/⚠️ Check 3: volume confirms - Volume is normal, but it is a pullback
- ✅/⚠️ Check 4: no major negative news - No news found
- ✅/⚠️ Check 5: healthy chip structure - Chip data unavailable
- ✅/⚠️ Check 6: reasonable PE valuation - PE is 20.276865, within reasonable range
 
### 🎯 Signal Attribution
 
**Attribution Weights**:
- 📈 Technical Indicators: 50%
- 📰 News Sentiment: 0%
- 📊 Fundamentals: 0%
- 🌐 Market Conditions: 50%
 
**🐂 Strongest Bullish Signal**: N/A
**🐻 Strongest Bearish Signal**: Strong bearish trend
 
### 💼 Financial Summary
 
| Report Date | Revenue | Net Profit (Parent) | Operating Cash Flow | ROE | Revenue YoY | Net Profit YoY | Gross Margin |
|:------:|-------:|-------:|-------:|------:|------:|------:|------:|
| 2026-07-31 | 962.21 亿美元 | 596.88 亿美元 | 240.77 亿美元 | 117.21% | 105.85% | 125.90% | 74.67% |
 
### 💵 Shareholder Return
 
| TTM Cash Dividend / Share (Pre-tax) | TTM Dividend Events | TTM Dividend Yield | Latest Ex-dividend Date |
|---------------------:|----------:|--------:|:--------:|
| 0.5200 美元 | 4 | 0.32% | 2026-09-10 |
 
### 🧩 Related Boards
 
Technology / Semiconductors
 
*📋 Data Sources:Technical data*
 
---
 
 
*Generated At:2026-11-13 09:37:32*
*Model:qwen2.5:7b*
======= reports/market_review_20261113.md =======
# 🎯 Market Review
 
## 2026-11-13 US Market Recap
 
### 1. Market Summary
The US market showed a cautious tone on 2026-11-13, with the major indices experiencing slight declines. The S&P 500 and Nasdaq led the downward trend, while the Dow Jones Industrial Average saw a milder drop. The VIX increased, indicating a rise in market volatility.
 
- **Market Signal**: 45/100 (mixed, balanced)
- **Drivers**: average major-index change -1.03%
- **Guidance**: Signals are mixed; keep position sizing moderate and wait for confirmation.
 
### 2. Index Commentary
- **S&P 500:** The S&P 500 closed at 7,513.96, down 1.03%, reflecting a risk-off sentiment.
- **Nasdaq:** The Nasdaq Composite closed at 24,946.52, down 1.65%, driven by tech-heavy sectors.
- **Dow Jones:** The Dow Jones Industrial Average ended at 52,090.60, down 0.39%, showing a more resilient performance compared to the other indices.
- **VIX:** The VIX, the fear gauge, closed at 20.37, up 8.93%, indicating increased market uncertainty.
 
| Index | Last | Change % | Open | High | Low | Amplitude | Turnover (USD bn) |
|-------|------|----------|------|------|-----|-----------|-----------------|
| 标普500指数 | 7513.96 | 🔴 -1.03% | 7616.85 | 7621.78 | 7499.40 | 1.61% | N/A |
| 纳斯达克综合指数 | 24946.52 | 🔴 -1.65% | 25408.00 | 25423.09 | 24888.98 | 2.11% | N/A |
| 道琼斯工业指数 | 52090.60 | 🔴 -0.39% | 52490.68 | 52521.48 | 52017.86 | 0.96% | N/A |
| VIX恐慌指数 | 20.37 | 🟢 +8.93% | 18.78 | 20.41 | 17.64 | 14.81% | N/A |
 
### 3. News Catalysts
No relevant news impacted the market on this day. The indices' moves were driven by internal dynamics and broader market sentiment rather than specific news events.
 
### 4. Outlook
The near-term outlook suggests a cautious stance, given the risk-off sentiment indicated by the S&P 500 and Nasdaq. The Dow's relative resilience suggests that the market may find support in more defensive sectors. However, the increasing VIX level suggests that volatility remains a key risk factor.
 
### 5. Risk Alerts
- **Increased Volatility:** The rise in the VIX indicates heightened market uncertainty. Investors should monitor this closely.
- **Key Levels:** Attention should be paid to key support and resistance levels in the major indices.
 
### 6. Strategy Plan
- **Stance:** Defensive
- **Position-Sizing:** Reduce overall exposure to high-risk assets and maintain a balanced portfolio.
- **Invalidation:** A significant breakout above recent resistance levels or a sustained decline below key support levels would invalidate this stance.
- **For reference only, not investment advice.**
Highlights
  • Runs free on GitHub Actions

    Copy the project to your GitHub account, add your keys as secrets, and GitHub's servers run it every weekday. No server to rent.

  • Six markets in one watchlist

    Covers China A-shares, Hong Kong, US, Japanese, Korean and Taiwan stocks plus ETFs, using free data sources like YFinance by default.

  • Chat with 15 built-in strategies

    A web workspace lets you ask follow-up questions using strategies like moving-average crossovers and Elliott wave, plus backtests and a portfolio view.

Star growth over time for ZhuLinsen/daily_stock_analysis
0xDipper (@Dipper_pol) lays out the five-step setup and ends with 'treat it as a second opinion, not a signal.'
Indie Hacker News asks whether this free GitHub stock analyst is actually useful, from setup to the daily dashboard.
Chapters — click to jump
The AI Index ranks it No. 3 in a top-10 GitHub trading tools roundup; starts at its section, which includes how to set it up.
NVIDIA · Business meaning for AI agents · Apache-2.0

NVIDIA builds a map of your business data for AI agents

ELI5

Imagine giving your AI assistant a glossary of your company's databases, so when you ask about revenue it knows which tables to use, shows its math, and tells you when the data cannot answer.

View on GitHub →
Stars
12.4k
Language
Python
Started
Apr 2026
GitHub repo card for NVIDIA/auto-ontology — click to open the repo
Highlights
  • Shows its work

    Every answer comes back with the database query and the steps that produced it, so you can check the number against the source.

  • Says when it can't answer

    An agent can first check whether your data covers a question. Asked about ad spend with no ad data, it asks you to clarify instead of guessing.

  • Plugs into Cursor and Claude Desktop

    Runs an MCP server (a standard plug-in port for AI tools) that Cursor and Claude Desktop can connect to, plus three ready-made agent skills.

Star growth over time for NVIDIA/auto-ontology
Gul Saeed Khattak (@Ldataengineer) explains the 'wrong revenue column' problem and lists the honest catches, including no published accuracy scores.
Sumanth (@Sumanth_077) walks through how it reconciles a metric like 'revenue' defined differently across two systems.
Syed Humair installs it, fixes three things that broke for him, builds the map from a sample database, and asks it questions.
Chapters — click to jump
Hacker News · 292 points, 139 comments
Developers say they can't tell what it does

On the Nov 3 Hacker News thread, one commenter wrote 'I've spent a career working with data and ontologies and I couldn't understand this readme,' and another asked why it beats long-standing standards like OWL/RDF and SPARQL. Another flagged that it needs an NVIDIA API key. The next day NVIDIA merged PR #404, 'Rewrite README around using Auto-Ontology and why it matters.'

Open the Hacker News thread →
google-deepmind · Robot physics, now version 4 · Apache-2.0

DeepMind's free robot physics simulator hits 4.0

ELI5

Think of it as a video game engine built for science: it calculates how robot arms, legs, cloth, and soft objects push, bend, and collide, so researchers can train robots on a computer before building real ones.

View on GitHub →
Stars
17.8k
Language
C++
Started
Aug 2021
GitHub repo card for google-deepmind/mujoco — click to open the repo
Highlights
  • Try it in your browser

    mujoco.org is now an interactive 4.0 showcase, and live.mujoco.org lets you drag in your own robot models with nothing to install.

  • Soft things behave like soft things

    Version 4.0 adds cloth that does not pass through itself, bendable rods, and plasticity, so simulated wire and clay keep their dents.

  • Ships with coding-agent skills

    The Python package now bundles the manual and agent skills so a coding agent can help build and debug simulations. A deformables skill landed this week.

Star growth over time for google-deepmind/mujoco
Lukas Ziegler (@lukas_m_ziegler) explains speculative contacts: why fast, thin objects used to pass straight through each other.
Boris Belousov (@_bbelousov) breaks down 4.0, including a robot arm catching a falling playing card, plus caveats and breaking changes.
Drift explains in about a minute what MuJoCo is and why robotics researchers use it.
Trossen Robotics walks through installing MuJoCo and running robot-arm simulations for its AI kit.
Chapters — click to jump
Kevin Wood trains a robot in MuJoCo and moves the learned skill onto real hardware, the sim-to-real story in one video.
Chapters — click to jump

My goal is to get to 100,000 subscribers by the end of this year.
Help me by subscribing.

73,000 today
0100,000
mattpocock · Engineering habits for agents · MIT

Matt Pocock's skills make your agent grill you first

ELI5

Before your AI coding helper writes a line, it interviews you about what you actually want, then builds in small tested steps. Think of it as a shelf of job recipes a careful senior engineer would follow.

View on GitHub →
Stars
323.5k
Language
Shell
Started
Feb 2026
GitHub repo card for mattpocock/skills — click to open the repo
mattpocock / skills

What's new

Three new commands in v1.3:

  • /implement-spec

    Implements an entire spec in a single run by reading tickets as a task graph, dispatching parallel subagents into isolated git worktrees, and merging them onto a single integration branch.

  • /pr

    Formats pull request bodies with a minimal visual summary (like pseudocode, a call tree, or Mermaid diagram), before/after proof, and a "merge-danger" blast-radius rating.

  • /retro

    Reviews a completed coding session and suggests improvements to your agent's environment (adding linter rules, pre-commit hooks, or CODING_STANDARDS.md updates) rather than the feature code itself.

Highlights
  • /grill-me interviews you first

    The agent asks question after question about your plan until every open choice is settled, before any code gets written.

  • A shared glossary for your project

    /grill-with-docs keeps a GLOSSARY.md of your project's own terms, so the agent stops using 20 words where one will do.

  • Self-updating in Claude Code and Codex

    Installs as a plugin that updates itself in Claude Code, Codex and GitHub Copilot, or as editable files you own and tweak.

Star growth over time for mattpocock/skills
Matt Pocock (@mattpocockuk) shows his 'Home Repo' trick: one repo that points agents at mattpocock/skills and his other projects.
Matt Pocock walks through v1.3: /implement-spec, /pr and /retro.
Chapters — click to jump
addyosmani · A senior engineer's playbook · MIT

Addy Osmani packs 25 engineering workflows for AI agents

ELI5

Give your AI coding helper the habits of a senior engineer: write a spec, plan small tasks, test, review and ship, each one a ready-made workflow with checkpoints it has to pass before moving on.

View on GitHub →
Stars
117.0k
Language
JavaScript
Started
Feb 2026
GitHub repo card for addyosmani/agent-skills — click to open the repo
addyosmani / agent-skills

What's unique

Most of his skills have a special section that rebuts the agent's rationalization: the excuses an agent uses to skip a step, each paired with a short reality check.

  • test-driven-development

    One rationalization in this skill is: "I'll write tests after the code works." The reality, from the repo: "You won't. And tests written after the fact test implementation, not behavior."

Highlights
  • Six commands from idea to launch

    /spec, /plan, /build, /test, /review and /ship walk the agent through each stage, and the matching workflows switch on by themselves.

  • /build auto runs the whole plan

    Approve the plan once and the agent builds every task, still testing and saving each one, pausing on failures or risky steps.

  • Google's engineering rules baked in

    Draws on Software Engineering at Google: changes of about 100 lines, the test pyramid, and treating old code as a liability.

Star growth over time for addyosmani/agent-skills
Yarchi (@undefinedKi) breaks down how each of the 25 skills lists the excuses agents use to skip steps, with a rebuttal for each.
Mario Nawfal (@RoundtableSpace) lists the nine commands and the '/build auto' approve-once mode.
Rob Shocks (Switch Dimension) explains Addy's six-stage lifecycle, then compares it with Pocock's pack and pstack.
Chapters — click to jump
The Gray Cat runs Addy's pack and Matt Pocock's on the same feature: both cost about $87, spent very differently.
Chapters — click to jump
AICodeKing walks through Addy's skills and wires them into the Verdent coding app.
Chapters — click to jump
cloudflare · Security audits for coding agents · MIT

Cloudflare's skill makes your agent a security auditor

ELI5

Ask your coding agent to security-check a project and it splits the job among separate helpers: some map the code, some hunt for holes, and fresh ones try to disprove each finding before it reaches your report.

View on GitHub →
Stars
71.6k
Language
JavaScript
Started
Jun 2026
GitHub repo card for cloudflare/security-audit-skill — click to open the repo
Highlights
  • Six phases, from map to report

    Maps where outsiders can reach your code, hunts by attack type, tests each lead, then writes a plain report plus a machine-readable findings file.

  • Every finding gets a skeptic

    The agent that checks a finding is never the one that found it. Leads it cannot prove stay marked needs_validation, with no severity.

  • Seeded Cloudflare's own scanner

    Cloudflare says this skill was the single-repo starting point for the fleet-wide vulnerability harness it described on its blog in June.

Star growth over time for cloudflare/security-audit-skill
TechToby (@techtoby__) ran it on his own code: 11 issues, 8 confirmed, but the deep setting ate his whole 5-hour Claude limit.
Mario Nawfal (@RoundtableSpace) explains the six phases and the 'fresh agent confirms every finding' rule.
Eric Tech runs the skill on his six-month-old app with real users: 26 problems, 21 real, then fixes them.
Chapters — click to jump
Rob Shocks plants 6 bugs in a SaaS app and scores free scanners, Claude Code's /security-review and this skill (round 3 starts here).
Chapters — click to jump
Signal Coders walks phase by phase through how a suspicious pattern becomes a confirmed finding or gets rejected.
GitHub Issue · cloudflare/security-audit-skill #128
An agent following the skill wiped a user's real home folder

On Windows 11 with no sandbox, a user says an agent running the skill tried to create the required 'scratch-local HOME' by clearing the real user profile, deleting ~/.claude (all conversation history), ~/.ssh, 128 installed skills and gutting the Python install. The reporter asks for a hard fail-closed check; a commenter adds that a sandbox-less audit burned about 43 agent calls and produced nothing confirmable.

Open issue #128 on GitHub →
⭐

First Stars

FROM THE AUDIENCE

Know which prompt shipped which line of code.

WHAT IT IS

Atlas records the sessions your team runs in coding agents like Claude Code and Codex and links each one to the commits it produced. When something breaks, you can see whose agent changed it and what it was told. It comes as a desktop app or a terminal tool, and works offline with no account in local mode.

View on GitHub →
Stars
9.4k
Language
TypeScript
Started
May 2026
Sent in by Andrew

Sent in through the team inbox. When a team ships with coding agents and something breaks, nobody can easily tell whose agent changed what, or why. Atlas is built to answer that.

Your turn

Building something? Send it to hi@TheNextNewThing.ai

FROM THE AUDIENCE

Talk to your coding agent.

WHAT IT IS

Bomby is a Mac app you just talk to. Tell it what you want and it hands the work to the coding agent you already use (Claude Code, Codex, OpenCode, goose or omp), chats with you while it works, and lets you know when it's done. It is MIT licensed and runs on your own keys.

View on GitHub →
Stars
1
Language
TypeScript
Started
Nov 2026
Sent in by Ahmed

Ahmed Aburady wrote in on Nov 9 and open-sourced Bomby after Andrew featured his imagine CLI on the show in August. Andrew, Nov 12: "add this to the next repo show I do. It's incredible."

Your turn

Building something? Send it to hi@TheNextNewThing.ai

FROM THE AUDIENCE

Team projects on GitHub, with chat built in.

WHAT IT IS

A small project-management app with native chat for teamwork. There is no cost and no backend: it uses GitHub Projects. If your team is already on Slack, you can plug that in instead of the native chat. When work gets done on GitHub, the issue moves in the app.

View on GitHub →
Stars
2
Language
Swift
Started
Nov 2026
Sent in by Oddmund

Oddmund wrote in asking for the First Stars segment. Krishna forwarded it on Nov 9, and Andrew asked Sage to add it to the markdown doc.

Your turn

Building something? Send it to hi@TheNextNewThing.ai

FROM THE AUDIENCE

The hack for mastering AI tools

WHAT IT IS

A Claude Code plugin that picks the model and effort for every prompt. It sends each prompt to the cheapest model and effort that can do the job, and raises effort only when Claude is actually stuck. It is MIT licensed, works with a Claude subscription or an API key, and installs as a plugin from archid21/rightsize.

View on GitHub →
Stars
2
Language
TypeScript
Started
Oct 2026
Sent in by Archi

Archi Dubey submitted rightsize and asked if it could go in this week's show rundown. Andrew: "Could you please include this in the show rundown? I'd love to talk about it if we get a chance this week."

Your turn

Building something? Send it to hi@TheNextNewThing.ai

My goal is to get to 100,000 subscribers by the end of this year.
Help me by subscribing.

73,000 today
0100,000