ChocoluvH
Hi HN! I'm really excited to share this. I'm the creator of Skill Compose.
I've been an NLP engineer for almost 10 years, and I've been a workflow fan ever since Google Dialogflow.
THE WORKFLOW RENAISSANCE (AND WHY IT MADE SENSE IN 2023)
Tools like Dify, and later n8n pushed "agent building with workflows". In 2023, that made sense for two reasons:
1) LLMs were unstable at basic tasks, so you needed deterministic nodes for checks and branching.
2) Instruction following was weak (especially open source), so workflows enforced control flow.
Workflows exploded. For a while, they were the right tool.
BUT SOMETHING DIDN'T FEEL RIGHT
Over the last two years, LLMs have improved at a pace that still shocks me.
But workflows haven't.
It feels like we're building agents in 2026 with a technique from 2016.
As logic grows, workflows get expensive to build, hard to read, painful to maintain, and awful to migrate.
We called it "Artificial Intelligence", but really?
Artificial, yes. Intelligence, not so much.
WHY I THINK SKILLS WIN (EVEN IF YOU STILL LIKE WORKFLOWS)
1) Skills enforce best practices.
LLMs are smart, but they can't reliably infer how your company works, like your slide standards or your security boundaries.
A skill is where you encode that reality as a versioned, reviewable, executable spec.
2) Skills evolve by local edits, not orchestration rewires.
Most iteration is about raising the standard, not reshaping the process. With skills, you improve rubrics, examples, specs, and deterministic helpers.
3) Skills don't collapse when reality drifts off the "happy path"
Workflows handle the happy path. When inputs drift or something breaks, they need a growing tree of edge cases.
Skills are different. Modern LLMs can write, run, and revise code to diagnose failures and recover, so the system can self heal instead of dead ending.
SO, WHAT SKILL COMPOSE IS
Skill Compose is an open source platform for building and shipping skill powered agents, no workflow graphs, and no CLI.
THE REAL GOAL: SKILLS FOR EVERYONE
Skill-powered agents should be accessible to anyone who has a repeatable process worth encoding.
---
Design Principle #1: Container first runtimes + Executors
Skill Compose is container first by default: agents run in containers (or K8s pods), not on your host.
You can run locally, but once an LLM executes code you start worrying about deps, random binaries, and breaking your system. So each agent gets its own isolated runtime for safer, reproducible execution.
By default, agents use the standard Skill Compose image. For heavier needs (GPU, large ML stacks, HPC), you can assign an Executor: a custom Docker image or K8s pod, even if it's tens of GB with CUDA, frameworks, and domain dependencies.
---
Design Principle #2: Evolution via feedback + traces.
Often times we wanted to give skills explicit feedback + real run traces to propose upgrades prompts, examples, scripts.
Start mediocre. Improve fast.
---
Design Principle #3: Skill management.
How about importing skills directly from a GitHub repo URL?
Also support portable formats: .skill files, .zip archives, and drag and drop folders.
Then treat skills like real software: version them, review them, diff them, and fork them.
---
Design Principle #4: No "skill marketplace" (AND WHY)
Black box skills you do not own are rarely useful.
The skills that matter are the ones you build, adapt, and fully control. They encode your workflows, policies, data shapes, edge cases, and permission boundaries.
Ownership is what makes skills auditable, secure, and maintainable.
---
THE HONEST DISCLAIMER
This is still early. Some parts are rough. Some parts are toy like. A lot will change.
---
Thanks! And please be blunt. Critical feedback is the fastest way this gets better.
Free AI Video Generator for Ghibli-Style Animations
1. Use ChatGPT-4o, upload a picture (for me I had a Titanic poster, you know..) with prompt "convert this photo to studio ghibli style anime."
2. Visit Movii.ai, log in with Google, upload the generated image as the first frame, pick a model (e.g. Wan/Vidu), choose 720p resolution, and set the duration to 5 seconds. I generated over 10 videos, totally free, no cc needed
3. Wait about a minute and voilà!
Check out what I got! (was very excited about this!!) =]
https://movii1s1.mai-cdn.com/projectt/video/655635988205062_...
I have tried AI tools like Freepik and Pollo.ai in the last couple months, Movii.ai is very new but feels cleaner and has no hidden tricks.
LGTM so far.
Try Milvus standalone instead, much simpler. I also just found their python version (https://github.com/milvus-io/embd-milvus), which is quite neat.
If your guy couldn't get a single open source software straight, you had the wrong guy :(
I can only see managed service useful when I had 100X traffic and when strong SLA is required.
The thing is to make ElasticSearch scores "comparable" to Milvus scores. Lots of ways to do this, but there's no single good solution. For example you could calculate BM25 score offline, or use TF-IDF score to do some kind of filtering. Again there's no single perfect answer. You'd have to do a lot of experiment according to your own use case and your own data to get the best results.
Also a lot of tuning needs to be done during all phases: 1) query pre-processing 2) query tokenizing 3) retrieval 4) ranking and reranking
I personally would not trust any universal "hybird-search" solutions. All toy demos.
It usually takes 5-10 good engineers to build a decent search engine/system for any real use case. It also requires a lot of turning, tricks, hand-written rules to make things work.
The real focus should be to improve the recall of vector search. Pity that nobody is doing real AI research here. Money wasted in marketing and branding.
My point is I only trust stuff that focuses their own business. Especially for small startups.
I am currently running with Milvus + ElasticSearch, works perfect. The latest Milvus version is super fast and scalable (>50M vectors). Haven't tried Zilliz Cloud. Have to find out what the cost is.
I am old school. IMO ElasticSearch is only good for keyword search and these so called "vector databases" products are only good for vector search.