Claude Fable 5 is now integrated with Claude Tag.

CC’s founder, Boris Cherny, and CC Product Manager Cat Wu recently sat down for an in-depth conversation, sharing the evolution from Claude Code to Claude Tag and how this technology has spread from the engineering team to the entire company.

For Anthropic employees, it’s already hard to imagine what a channel without Claude would even look like.

Andrej Karpathy has called Claude Tag a paradigm shift in LLM interaction. You can read my previous article on this:

AI Disruption

Anthropic Releases Claude Tag

“AI Disruption” Publication 10100 Subscriptions 20% Discount Offer Link…

Read more

9 days ago · 11 likes · Meng Li

Triple Jump: From Autocompleting a Single Line of Code to AI Scheduling Its Own Work

Three years ago, Anthropic had just begun introducing AI into daily workflows. Two years ago, employees were still using code autocompletion—AI could only complete the current line being written, while the human still decided what to write. Then Claude Code launched, bringing agents. AI became capable of writing complete functions, entire files, and even full features. Now, with Claude Tag, AI can independently complete entire tasks—implementing features, running experiments, performing data analysis—handling the full end-to-end process on its own.

Read more

A creator named @reidhannaford used Blender to build a simple gray-box scene — a few cubes representing buildings and a cylinder representing the character — then created a 5-second camera movement: the shot slowly pushes in from a distance, orbits around the character, and finally settles on a close-up.

The entire scene looked extremely rudimentary, like early-game footage.

But when he fed this Blender-rendered reference video into Seedance 2.0, something magical happened.

Seedance perfectly replicated the camera trajectory, with the character’s movement through space matching frame by frame. Even more astonishingly, the originally crude gray-box scene was transformed into a richly detailed, realistic world — the cubes became towering skyscrapers, the cylinder became a real person, and the lighting, textures, and atmosphere all reached cinematic quality.

This video reminded me of how, just six months ago, I was still struggling with the “out-of-control” feeling in AI video generation. Back then, I kept wondering: Can AI video generation ever achieve the precise control of traditional filmmaking? Or are we doomed to only hunt for occasional happy accidents within AI’s randomness?

After watching this video, I finally realized: Yes, it can.

Read more

A couple of days ago, I came across an open-source project for automatic video editing. You just say to the AI: “Turn all this into a publishable video,” and it automatically removes filler words, adjusts colors, adds subtitles, and even creates animations.

It’s an open-source project called video-use. It hit the GitHub daily trending TOP 5 yesterday and has already earned 12,000 stars.

Anyone who creates content knows that editing is the most tedious part. The learning curve is steep, and it’s hard for ordinary people to get started. Even relatively simple tools like CapCut still require you to drag timelines and tweak parameters — you’re essentially “operating software” rather than “expressing creativity.”

That was until I discovered video-use.

Read more

Just saw this—Firecrawl’s official account posted it.

Starting today, you can use Firecrawl without applying for a Key or setting up any env variables — just call the API directly.

They’re also giving away 1,000 free credits per month.

Firecrawl is a web data API built specifically for AI.

It can turn any webpage into clean Markdown or structured data that AI can read directly.

Give it a URL, and it returns:

• Super clean body Markdown (removes navigation, ads, footers, and all the noise)

• Or structured JSON (you define the schema, it extracts accordingly)

• Screenshots, full HTML, metadata — whatever you need.

It also supports full website crawling, local file parsing, arXiv semantic search, asynchronous browser research Agents, GitHub repo info lookup, and more.

Read more

Recently, an open-source project has been consistently trending on GitHub, currently sitting at 44.1K stars.

It’s a tool called Agent-Reach.

Today’s AI Agents are already incredibly powerful. Claude Code can refactor entire projects, OpenClaw can handle daily tasks, and Cursor can complete full-stack development right inside the IDE.

But there’s one thing they’ve never been good at: fetching information from the internet.

Ask your Agent to research a technical solution:

• “Go find similar open-source projects on GitHub” — authentication setup gets stuck for ages.

• “Check community feedback on Reddit” — 403 error, anonymous APIs are blocked.

• “Search Twitter for user reviews of this product” — Twitter API is paywalled, even the cheapest tier costs hundreds of dollars.

Every platform has its own barriers. You end up stepping through pitfalls one by one — installing tools, tweaking configs. Just getting an Agent to read a tweet can take half a day of troubleshooting.

Read more

In the field of brand design, logo motion effects are no longer just a nice-to-have addition — they have become a core component of brand experience.

The reality, however, is harsh: professional motion designers are scarce and expensive, ordinary designers often lack motion development skills, and developers typically don’t understand design aesthetics.

This has led to an awkward situation — many excellent design drafts can only be presented in static form.

Just when everyone thought creating logo animations was destined to be a painful chore, an open-source project called pixel2motion emerged and completely changed the game.

Read more

In recent months, I’ve found myself trapped in what I call “Terminal Hell.”

After Claude Code exploded in popularity, every major AI company rushed to release its own Code CLI tool. As a result, my desktop now looks like this: three terminal windows side by side, running Claude Code, Codex, and Gemini respectively.

Whenever a new requirement comes in, I turn into an air traffic controller: I send a command to Claude Code, switch to Codex to check progress, jump back to Claude Code to review the output, and finally hop over to Gemini to wrap things up.

This workflow has a fatal flaw — the context-switching cost is way too high.

I often forget what the previous Agent was doing while switching terminals, or I mix up which branch belongs to which task. Even worse, sometimes two Agents end up editing the same file, causing merge conflicts that I then have to spend time resolving.

So, is there a way to let multiple AI coding assistants work in parallel while keeping their outputs centrally managed?

That’s when I discovered Orca.

Read more

Recently, while working on a RAG project, I needed to feed a large volume of technical documentation into a large language model.

At first, I used a conventional web crawling approach, but the retrieved content was disastrous—navigation bars, footers, ads, and script code all mixed together. In a 50,000-token HTML page, the truly useful content was less than 800 tokens.

Even more frustrating was that some websites had Cloudflare protection, directly returning 403 errors or captcha pages. Using Playwright to run a headless browser was slow and heavy, with each request taking several seconds.

I searched through countless projects on GitHub and tried tools like trafilatura, newspaper3k, readability, etc., but the results were all unsatisfactory.

Until I discovered webclaw, it truly opened the door to a new world!

Read more

Tired of accumulating massive amounts of notes in Obsidian but struggling to organize them efficiently?

Faced with hundreds or even thousands of documents, do you still have to manually dig through them just to extract key points and summarize content?

As a knowledge management enthusiast who works with Obsidian every day, my vault has ballooned to hundreds of notes. Every time I need to pull useful information from it, it feels like searching for a needle in a haystack.

That was until I recently discovered Claudian — an absolute game-changer.

Read more

Today’s smart homes have a fundamental problem—not that they aren’t “smart enough,” but that the approach to intelligence is wrong.

We’ve spent far too much energy on “making devices controllable” instead of “making devices think proactively.”

As a result, every piece of hardware sits idle, waiting for a command—from your phone, a voice assistant, or a preset automation rule. And once your daily rhythm changes, those rules all break.

What if home devices could understand when you need something and act on their own?

Today I recommend an open-source project: Anima. It delivers exactly what you’re looking for.

Read more

The black hole’s appetite is exactly equal to Claude Code’s context window.

When you’re chatting with Claude Code, and the context hits the limit, the model quietly runs /compact on its own — and suddenly the details from the first half of your project get blurry.

It always feels jarring.

There’s an open-source project that turned this into an animation. While using Claude Code in the terminal, the fuller the context, the bigger the black hole gets.

Read more

If you’ve ever asked an AI Agent to create a PPT, you’ve probably experienced that awkward moment of “I wrote it, but I can’t see it.”

For a long time in the field of document automation, both developers and Agents have been relying on the same set of Python libraries (python-pptx, python-docx, openpyxl).

These libraries can read and write structure, but rendering belongs to an entirely different world. Want a preview? You have to install LibreOffice or call Microsoft’s Graph API. The Agent runs smoothly on the server—until it hits the rendering stage and gets stuck.

This leads to a very embarrassing situation: Agents can “write” documents, but they cannot “see” them. They have no way to self-inspect what they’ve created, and delivery quality is left entirely to luck.

Until OfficeCLI arrived, and things finally changed.

Read more

Can AI generate a complete 2D game character animation spritesheet for you using just one reference image and a few natural language descriptions?

Not a “looks like animation” video, but actual PNG spritesheets that can be directly imported into a game engine — every frame is a properly redrawn independent pose, not stretched or deformed “fake animation” from the original image.

A new project has recently appeared on GitHub that turns this into a visual workspace running directly in the browser.

Project in One Sentence
Gorest 2D Animation Spritesheet Generator (hereafter referred to as Gorest) is a local browser-based workspace for building side-view 2D game scenes, importing/generating spritesheets, previewing animation clips, and organizing reusable scene assets.

Its core concept can be summed up in one phrase: “Codex-Assisted Animation Agent”.

Read more

Over 60% of users get stuck at the “installation and configuration” stage on their first attempt with AI Agent tools.
Before they even get to experience the magic of AI programming, they’re already discouraged and give up.

This isn’t the tool’s fault, nor the user’s. Network issues, environment setup, configuration, and authentication — every single step is full of pitfalls, and the pitfalls affect each other. One wrong move and everything downstream falls apart.

So the question arises: Is there a tool that wraps all these most troublesome parts into a single graphical interface, allowing users to get tools like Claude Code, Codex, OpenClaw, Aider, and others up and running with just a few mouse clicks?

The answer is yes. It’s called EchoBird.

Read more

Google today released an open specification called Open Knowledge Format (OKF).

The problem it aims to solve is one that almost every team building AI agents has encountered: models are getting more powerful, but the contextual knowledge they need is scattered everywhere, unorganized, and lacking a universal format.

Open source repository:
https://github.com/GoogleCloudPlatform/knowledge-catalog/tree/main/okf

The stronger AI becomes, the more deadly knowledge fragmentation is

In most companies, the knowledge AI agents need is internal: the meaning of fields in a data table, business definitions of metrics, system access paths, API deprecation notices, and so on.

Where does all this live today?
Everywhere. Metadata catalogs, internal wikis, shared documents, code comments, and the heads of senior engineers. Each system has its own API, and each platform has its own data model. They are incompatible and don’t interoperate.

The result: every team building agent ends up solving the same problem from scratch — how to piece together these scattered pieces of knowledge and feed them to the model. Every vendor reinvents the same data structures, while the knowledge itself remains locked inside the platform that created it.

Read more

A couple of days ago, while scrolling on X (Twitter), I saw many creators making AI short dramas frantically complaining: when using mainstream video models to generate clips with “dangerous actions” or “subtle breathing sounds,” the platforms kept rejecting them. They revised the prompts over a dozen times, but still triggered the review filters every time.

This has become all too common in the AI image and video space this year. To avoid compliance risks, commercial platforms have set their review thresholds increasingly aggressively. Top-tier models like Sora 2, Veo 3.1, Kling 3.0, and Seedance 2.0 are excellent, but if you want to create something creative or with an edgy style, you almost have to fight a battle of wits with the censors.

Even worse is the subscription cost. Each platform has its own accounts, payment systems, and exclusive models. Heavy users can easily spend a fortune on subscriptions every month. Yet all of that can be replaced by a single open-source project.

Read more

What It’s Like to Job Hunt in the AI Industry

You open a job board, type “AI Engineer,” and 200+ positions pop up. You click through each one, read 800-word job descriptions, and narrow it down to maybe 20 that look promising.

Then you spend another two hours tailoring your resume, writing personalized cover letters, and filling out 15-field application forms. Two weeks later, you receive 18 automated rejections that all start with “After careful review, we have decided...”

This isn’t one person’s story — it’s the daily reality for every technical job seeker right now. Companies use AI to screen you, while you manually screen companies.

Santiago Fernández de Valderrama (known as Santifer) went through this painful process. After months of grinding the hardest way possible, he decided to fight fire with fire — he built his own AI system to “reverse-screen” companies.

Using this system, he evaluated over 740 positions, generated 100+ tailored resumes, and ultimately landed a Head of Applied AI role.

Even more impressively, after he open-sourced the system, it exploded to 53.5K+ stars on GitHub and was covered by Wired, Business Insider, and other major outlets.

This system is called Career-Ops.

Read more

Recently, the BrowserAct team open-sourced a set of Skills specifically designed to solve a common problem faced by AI Agents: being frequently blocked by various anti-bot mechanisms (such as CAPTCHA and human verification) when automating browser operations or scraping data.

This Skill package connects these challenges into a complete, end-to-end pipeline.

Project Overview

BrowserAct Skills is an open-source AI Agent browser skill toolkit developed by the BrowserAct team under Singapore-based ECOCREATE TECHNOLOGY PTE. LTD. It focuses on empowering various large language model (LLM) AI agents with real browser control capabilities.

The project has already surpassed 2.4K stars on GitHub and is growing rapidly.

The project is divided into two independently installable core components:

• browser-act: Real-time browser execution module that acts as the “hands and feet” of AI Agents. It comes with 50+ standardized browser operation commands.

• browser-act-skill-forge: Skill factory generation module designed for recurring, periodic web tasks. It enables “one-time debugging, permanent reuse”.

Read more

All independent developers know this: writing the code is just the beginning. The real headache is deployment.

Cloud servers, configuring Nginx, applying for SSL certificates, binding domains, uploading files… If you also need to add a backend and database, the whole process can easily eat up half a day.

If you just want to quickly validate an MVP or show a prototype to a client, the effort-to-reward ratio for all this infrastructure work is simply too low.

This is the problem PinMe aims to solve. Its core promise can be summed up in one sentence:
Create and deploy your web in one command.

One single command handles the entire flow from project creation to live deployment.

Read more

Recently, I came across a refreshing project on GitHub called harness-anything — essentially a connector that bridges AI with desktop software.

It allows large language models to directly control local applications on your computer, such as WPS, Photoshop, Illustrator, and more.

The AI can take over these programs just like a colleague who knows how to use a mouse and keyboard.

Project Introduction

harness-anything is a standardized command-line toolkit that currently provides 47 commands, covering three major scenarios: office work, design, and academia.

Read more