After Completing the Nuwa and Darwin Skills, I Found the Missing Piece of My Skill Puzzle: MetaSkill

Over the past two months, I’ve been diving deep into Agent skills — partly to standardize and abstract my daily workflows into skills, partly to figure out how to let AI build skills on its own.

First came Nuwa. It can distill anyone’s thought process into a runnable Skill. Then came Darwin, which lets a finished Skill evolve on its own — iterating lower-scoring versions to higher ones automatically. One handles “birth,” the other handles “improvement.”

Two puzzle pieces in hand. But the only problem? My skill collection was growing chaotic.

When you ask an agent to do something complex, it needs to pick, combine, and sequence multiple skills. Even now, I often have to trigger and configure that process manually. Setting up a writing workflow means I have to think through: first search for material, then learn the style, then write a draft, then proofread — mapping it all out step by step in my mind. No matter how many skills I build or how well I optimize them, the third puzzle piece — getting the agent to organize them on its own — remained unsolved.

First, What Is OpenSquilla?

OpenSquilla is an open-source, locally runnable AI agent written in Python. Think of it as competing in the same arena as OpenClaw and Hermes Agent, but it’s taken a few more steps down the open-source agent path, with more advanced strategies for model invocation and skill development/management.

Its product philosophy boils down to one sentence: Don’t compete on models; compete on the Harness.

“Harness” is the shell outside the model. The model is the engine; the harness is the transmission, throttle, and navigation system. While everyone else races to increase horsepower, OpenSquilla optimizes the shell: which model to call at each step, which skill to use, and how to avoid burning tokens unnecessarily.

This harness is built in two phases: v1.0 integrates intelligent model routing into the base, and v3.0 adds MetaSkill on top. I focused on testing these two capabilities, and also took the routing base for a spin.

First Thing: It Recognized My Existing Skill Library Right Out of the Box

When I installed it and opened the skill list page, I was pleasantly surprised.

It recognized 134 skills. Besides its built-in ones, it picked up all my personal skills from Claude Code: Feishu full-suite, browser automation, my own Nuwa, Darwin, topic generation, table processing — all appeared in the list exactly as they were.

I didn’t import anything. It scanned them automatically.

The skill ecosystem is exploding right now, with thousands of Skills popping up in the community — some human-written, some AI-generated. But what each person actually uses effectively is the small set they’ve accumulated. OpenSquilla didn’t force me to reconfigure from scratch; my existing skills were ready to use.

Second Thing: I Told It to Create a New Skill with One Sentence, and It Actually Did

This was what I really wanted to see.

It has something called MetaSkill Creator. It claims that by describing a need in natural language, it can automatically synthesize a new skill — compressing the half-hour it takes to hand-write a skill down to a few minutes.

I didn’t hold back. I threw in a request:

“First fact-check a Chinese draft, then rewrite it into a more conversational, AI-free version, and finally output a list of modifications.”

It ran for a bit and generated a skill file. Not just a description — a complete flow chart, like this:

fact_check (事实核查) → rewrite (去AI腔改写) → modlist (生成修改清单)

It automatically filled in which skill to use at each step, which previous step it depends on, and where to pass the results. I’ve run this multi-step setup based on a flow chart, and it works.

I forgot to record the polishing skill test earlier 😓. So to give you a complete, uncut demo, I recorded another request from scratch: asking it to create a skill that can give editing suggestions based on video subtitles and even write a script. Watch it piece things together step by step:

This is exactly the puzzle piece I was missing. Nuwa — my creation — lets AI build individual skills. Darwin — also my creation — lets individual skills improve. The third piece — organizing multiple skills into a real, runnable pipeline — I had never built. This time, OpenSquilla’s MetaSkill filled that gap for me. Creation, optimization, organization — now it’s finally complete.

So What Exactly Is MetaSkill?

I think their official definition is a bit convoluted. Let me put it in plain English.

Essentially, it’s a special Markdown file. At the top it says “this is a meta skill,” and below it connects several steps into a flow chart. But the key isn’t the file itself.

Think of it this way. In the old orchestration frameworks, I hard-code the workflow in advance, and the agent runs along the lines I drew. MetaSkill flips that: I don’t write the workflow; I just tell the model the rules and available skills, and let it figure out its own workflow on the fly. One is me drawing a route map for you; the other is me giving you a map and rules, and you find your own path.

You describe the goal in natural language, and it handles the rest: selecting relevant skills, composing them into a workflow, and arranging tools and context.

And it doesn’t let the agent run wild. Once the flow chart is handed to the runtime, dependencies, tool whitelists, risk levels, and template security are all enforced. Which steps can read files, which can run commands — all must be declared in advance. It lets the agent organize itself, but with guardrails in place.

In one sentence: Before, I had to figure out how to compose skills and feed that to the agent. Now, I just state the goal, and the agent composes the skills itself.

Bonus: A Money-Saving Insight

During testing, there’s another thing worth highlighting: its model routing — the v1.0 base I mentioned earlier.

I first tested it with Volcano Engine’s Doubao (豆包). OpenSquilla has four built-in model tiers, from the cheapest “mini” to the most capable “code” version. The router evaluates the difficulty of each task and selects automatically.

I asked it to explain HTTP in one sentence. It picked the cheapest mini tier, no thinking mode. Then I gave it a harder problem: wolf, sheep, cabbage crossing the river, plus analyzing the abstract commonality with graph coloring. It upgraded itself to the strongest tier and turned on deep thinking. The input price difference between these two tiers is about 16x.

Limiting to one provider wouldn’t be fair, so I switched to OpenRouter and tested again. It now picks across dozens of providers’ models. For the same HTTP question, it routed to DeepSeek’s cheap flash model. For the tough river crossing problem, it jumped straight to Claude’s Opus with deep thinking enabled. Simple jobs go to cheap Chinese models; hard jobs go to the most expensive flagship — and this selection is made across providers.

Another plus for domestic users: It’s not picky. Besides overseas aggregators like OpenRouter, it can also connect directly to major Chinese cloud providers like Volcano Engine, Alibaba Cloud, and Tencent Cloud. You can go fully domestic and it still runs fine.

No matter which approach, the savings are the same: use cheap when possible, only spend on the truly difficult rounds. And that difficulty judgment is done locally on my machine — it doesn’t send my question to an external model for scoring.

A Deeper Observation

Currently, almost every agent product throws model selection in the user’s face: a dropdown menu of models, pick one. I’m good at picking — I do model evaluations daily, know who’s good at coding, who’s good at reasoning, who’s cheap.

But precisely because I do this every day, I increasingly believe this isn’t the final form. Why should ordinary users have to know all this? Making users worry about which model to pick for every conversation is a failure of product design. Automatic routing removes that burden — you just say what you want done, and it handles the allocation.

This is the real meaning of “competing on the Harness.” With the same budget, you can get much more done.

This was just a quick demo of a single request. I also checked the total cost after running 25 tasks: compared to a plan using only Opus, the scores were virtually identical, but the cost was an order of magnitude lower.

Final Thoughts

Back to those three puzzle pieces.

Creation (Nuwa) and optimization (Darwin) — I built those myself. The third, organization, was always missing. “Puzzle piece” isn’t quite right — organization feels more like a layer on top of the first two. Without it, no matter how many or how good the skills are, they work in isolation and can’t be chained together.

My strongest takeaway from these two months is: When skills reach a certain scale, the bottleneck is no longer how smart the model is, but whether these skills can be organized. OpenSquilla’s answer is to build “organization” into the agent’s base layer — let the agent handle it, not wait for me to configure everything manually. From v1.0’s intelligent routing to v3.0’s MetaSkill, it’s all about “competing on the Harness, not the model.”

The industry’s attention is almost entirely on “the model’s score went up again.” But could workflow orchestration be becoming more important than parameter count? I can’t say for sure. But for someone like me who has a pile of skills, I truly hope this direction works out.

It’s open source. Search for OpenSquilla on GitHub and you’ll find it. If you want to try it yourself, go take a look. In my case, it really did fill that missing puzzle piece.