Jay from Not a Flat Temple | QbitAI Official Account
The latest hot repository on GitHub: OpenSquilla.
After Lobster and Birkin, another Agent is rapidly climbing the charts, now with over 2,000 stars.
A quick look reveals its biggest selling point is a built-in "Smart Model Router". This means, for the same task, the token cost can be 60-80% lower than Lobster's.
When running tasks, it looks like this, just like a "slot machine".
It's very satisfying. When a task is done, an animation pops up showing how many tokens you saved, and you can trace the token consumption for each step.
It often saves over 90%, making you feel like your wallet suddenly got fatter.
It's absolutely insane, a top-tier "miser" (just kidding).
But that's not what I want to talk about today.
They just released a new feature called Meta Skill.
Not the company Meta, but literally—a Meta Skill, a Skill for Skills.
A Meta Skill embeds multiple Skills within it. Stitched together, it forms a super white paper that can break down and execute an entire end-to-end long-chain workflow.
△ Image generated by AI
Meta Skill Hands-on Test
Coincidentally, I had this exact idea last-last week when automating the backend work for our AIGC Industry Summit.
You might not know, but organizing such a conference is very tedious, and the work chain is extremely fragmented.
I previously organized the SOP, and the part I was responsible for looked something like this.
It looks very clear, but to automate it, it means that each blue square in the picture would require manually calling a Skill in the dialog box.
It was like poking a toad to make it move once, with a human in the loop the entire time. Just scrolling through the Skill list was enough to keep you busy for a long time, which was extremely annoying.
So, after the workflow was solidified, I refined all this context into a super Skill, making it end-to-end.
Once it gets the context, it automatically determines which stage it's currently in and then calls the corresponding sub-Skill to deliver the result.
Coupled with a heartbeat mechanism to periodically check the status document, it can be fully automated.
But honestly, creating such a large package of Skills was really troublesome, and the end result was actually quite crude.
And this was under the premise that I myself have know-how in the content industry. If it involves a task that requires integrating expertise across different industries, arranging and combining Skills like this would be a genuine disaster.
So, I've been frantically searching for something pre-packaged, like Claude's built-in Skills.
(A shameless free-rider.)
Lo and behold, I discovered just the day before yesterday that such a repository actually exists on GitHub!
Nine pre-packaged Meta Skills, the kind embedded within the Agent.
Let's pick one for a real-world test. Let's choose the one called meta-kid-project-planner.
A Meta Skill for children's project planning. When a kid needs to do a science fair project, a hobby craft, or a creative little invention, it helps them formulate an executable project plan from scratch.
Mostly because June 1st, International Children's Day, is just around the corner. It felt quite fitting. To save some effort, I designed the scenario so the child would learn about Meta Skills themselves and help me explain what each of these 9 Skills does to everyone.
The prompt was like this:
A 9-year-old child wants to make a Meta Skill spellbook, first presented as a webpage, then made into a small paper book, with each page introducing a spell.
No need to write in great detail. Before officially starting, based on the Skill instructions, it will ask you to provide more information.
Then, it entered the "go completely bonkers" phase.
It's absurd. It instantly spat out a massive execution trace, decomposing the requirements into a long string of sub-tasks.
No human intervention is needed the entire time. It ran on its own for over 20 minutes and ultimately delivered a complete 7-day project planning package.
It was indeed quite comprehensive. Even considering that the user is a child, it performed an extra round of safety review.
It even prepared various contingency plans.
How did it achieve this?
I looked through the SKILL.md source file for this Meta Skill. The general workflow is like this:
1. Project Initiation: Ask for user preferences—age, duration, budget, level of parental involvement.
2. Feasibility Classification: Determine if it's safe, whether adult help is needed, whether extra purchases are required.
3. Execution: Step-by-step plan → materials checklist → safety reminders → parent learning objectives → final assembly and delivery.
If outdoor activities are involved, it will even call a web search to check the weather…
△ Image generated by AI
The entire workflow is stitched together from 5 different atomic Skills.
The final deliverable is the project planning package markdown file mentioned above, around 3,000 words long.
But to make it easier to view, I had it use Claude Opus 4.7 to compile this context into a HTML file.
The interaction logic is like a physical book, with a page-turning animation every time you flip a page.
All 9 Meta Skills are explained according to my "Harry Potter-style" instructions.
At the end, there's even a "Magic Mirror" function, making it easy for children to choose a Skill based on their needs.
The webpage has been deployed. The link is at the end of the article. If you don't want to read the markdown, you can understand these 9 Meta Skills through this format.
Right, there's one more rather important thing.
From my testing, it feels like this Squilla router indeed saves quite a bit of money, not letting Claude handle everything.
For instance, for a trivial question like this, it only cost me three cents.
(P.S.: DeepSeek is really a great deal after the price cut!)
Looking at this token cost benchmark, it's quite impactful.
Of course, you can also choose whether to turn it on, or directly prompt it to lock onto a specific model.
Finally, let's talk about installation.
To be honest, it's pretty simple. If you've installed Lobster, it supports one-click migration of data assets and API keys.
The annoying part is that Mac requires installing via the terminal. Currently, only Windows supports a compressed package.
But no worries. For Mac/Linux, just copy and paste this string of commands in the terminal in sequence. If you run into any problems later, just toss them to Codex.
# Install uv:
curl -LsSf https://astral.sh/uv/install.sh | sh
. "$HOME/.local/bin/env"
# Install OpenSquilla:
uv tool install --python 3.12 "opensquilla[recommended] @ https://github.com/opensquilla/opensquilla/releases/download/v0.3.0/opensquilla-0.3.0-py3-none-any.whl"
# Configure and run:
opensquilla onboard
opensquilla gateway runWhen this interface appears, the installation is successful.
In terms of entry points, it supports all major IM platforms—Feishu, Discord, QQ… Just choose the one you're most comfortable with.
But I still recommend using their web version. No special reason, except that the "Smart Router Slot Machine" and "Token Saving Animation" are only displayed on the web version…
If you use a terminal or IM, you'll miss out on quite a bit of the fun.
What's the Secret?
At this point, some might wonder: For that project planning package that ran for a full 20 minutes, calling search engines, checking the weather, looking up memory, doing safety reviews in the process… who was orchestrating all these tasks?
With previous Agents, one Skill did one thing. Search was for searching, documents were for documents, weather was for weather. You had to manually string the workflow together, constantly keeping in your head which Skill to call next.
But now, Meta Skill has taken on this "dispatcher" role.
You can think of it as a "project manager's operations manual"—which steps are parallel, which are serial, which step's output feeds into the next… it's all written inside.
But having a project manager alone isn't enough.
A PM who just schedules work, if they call the most expensive contractor for every single step, the company will go bankrupt before long.
That's why I kept mentioning the Smart Router. It truly silently bears so much of the burden.
The "slot machine" you saw earlier is helping this PM with budget management. As each sub-step passes through, the router tightly clamps down on your wallet for you.
For example, with the kid-project-planner just now, extracting the child's age and preferences could be handled by DeepSeek; only steps like generating a safety review plan and the 14-day plan needed to be assigned to a model with more parameters.
Finally, we have the project manager, and we have budget management. But there's still one problem—
Who writes the PM's operations manual?
Meta Skill is great, but the key point is, it's really complex!
Based on personal experience, that kid-project-planner just now, with over 400 lines in SKILL.md, even iterating with AI would take about 30 minutes. And this assumes you already have a clear SOP in your head.
So, a Meta Skill for creating Meta Skills (such a tongue twister) is also very necessary.
It's called meta-skill-creator. I think this is the most important Skill among the 9 Skills released by OpenSquilla this time.
I recommend everyone go read this md file in the repository. If you can understand this, you can basically grasp the working principle of the entire Meta Skill.
But if you don't have time, you can also directly look at this AI-generated explanatory diagram.
△ Image generated by AI
At this step, all the barriers to Meta Skill have been resolved.
But the team was also thinking about something farther down the line—
The problem of supply-demand matching.
As Creators continuously produce new Meta Skills and the community contributes more, how to solve the problem of Skill inflation?
There are only 9 in the repository now, but if there are hundreds of Meta Skills in the future, how do you know which one best fits your scenario?
The solution OpenSquilla offers is: an indexing protocol for "Individual × Community".
Which Skills you usually use, what combination order you prefer, which ones didn't work well when tried… these act as signals, matched by the Agent with Skills made by others in the community, and then stitched together into a new one according to your workflow.
Simply put, an automatic Skill recommendation engine.
The Skill 2.0 Era Has Arrived
Lastly, let's meticulously comb through this repository's version history. This perspective is quite interesting—like unraveling a cocoon, it gradually reveals their product thinking.
In early May, the shoe first dropped with the release of the Smart Router. At first, it seemed like it was purely for token optimization.
Now, with the release of Meta Skill, the footprint is finally becoming clear, and it points in a counter-intuitive direction.
It stands at the intersection of three lines—
1. Models.
The ability to understand complex, multi-step instructions has skyrocketed in the past two years. The Agent token data flywheel has started turning.
Models can now "understand" complex orchestration instructions. This is the prerequisite for everything.
2. Ecosystem.
Community-created Skills are growing explosively.
From users handwriting them, to automatic generation based on data, to community aggregation and sharing… When there are thousands of optional Skills, you need a higher abstraction layer—Meta Skill—to reduce the cost of sifting through Skills.
3. Cost.
Running large models at scale is still expensive. Every time you let an Agent do trial-and-error online, repeatedly exploring the optimal path, a massive number of tokens get burned.
With Meta Skill, you can directly solidify this layer of complexity, moving the optimization problem upstream to the Skill layer.
These three pain points simultaneously point toward another paradigm shift being forced into existence.
Skill 2.0.
Individual Skills are no longer sufficient. For automation to go deeper, it must learn to arrange and combine multiple sub-Skills.
The next problem for Agents to solve has shifted from "Can it call a tool?" to "Can it organize tools?".
But from another perspective, it's actually quite exciting—
What kind of imaginative space can be created when multiple Skills collide?
After all, recently, many models have launched their own Agent teams. Tencent has Marvis, MiniMax has Mavis, Kimi has an Agent cluster…
But the Skill layer seems to be stuck at the phase from when Claude first popularized it. The community is basically still writing SKILL.md for individual models.
So I feel, the potential of multi-Agent systems has never been fully unleashed.
And this time, the emergence of Meta Skill lets me see a possibility—a white paper specifically designed for Agent teams, endowing models with a more macro-level global context.
Agents and humans are actually converging on many core problems.
When the number of employees (Agents) increases, and the number of business lines (Skills) increases, you inevitably encounter exponentially amplified noise.
At that point, mastering the use of architecture and management for entropy reduction becomes absolutely necessary.
Welcome to opening Meta Skill.
This is the white paper used to guide the Agent's entire ministerial bureaucracy.
One More Thing
Oh, right! Wait, don't click away just yet.
I found a strange thing while researching—
For such a hot project, there's only an official account on X (Twitter), and the very first version was retweeted by the big shot AK, yet no one knew who the team was. They were quite mysterious.
Upon inquiring, I heard it was a Chinese team, so I became even more curious.
After fruitlessly digging with AI for half a day, I finally resorted to the ancient method of asking around. Unexpectedly, Wang Yunhe's latest entrepreneurial news was thus exposed:
OpenSquilla comes from a company he founded called "Jiyuan Lüdong" (Fundamental Rhythm). The team is actively hiring.
So it turns out the team just got started and wants to focus on polishing the product.
Let the bullet fly a little longer. We'll watch and see.
GitHub: https://github.com/opensquilla/opensquilla
Skill Spellbook: https://imtangyujing.github.io/opensquilla-meta-skill-grimoire/