jinfeng79
Joined 3 years ago
Karma
3
Submissions
25
Member since
Feb 16
Recent submissions
Comment · 9 months ago
Comment · 10 months ago
Author here.
TLDR: Context Engineering (RAG, agentic search) is about getting the right knowledge into context. That's mostly solved now. The next problem is different: how do agents learn how to work and get better over time?
Skills are the mechanism. Once a task becomes a Skill, you don't do it anymore—the agent does. You move up. When your new work becomes a Skill, cycle repeats. Both you and the agent keep compounding.
Curious what others are seeing.
Comment · 10 months ago
One more thing—AI-generated Skills are way more detailed and specific than human prompts.
Humans write vague stuff like "make it better." AI writes 50 lines of exact instructions. That specificity = stability.
So ironically, AI writes better instructions for AI than humans do.
Comment · 10 months ago
Good question.
A Skill is basically a markdown file with instructions + maybe some scripts. You can literally read what it tells the AI to do. Line by line.
So when something goes wrong, you open the file, see exactly what instruction caused it, fix it. Version control like any code.
Compare to raw agent—you give it a prompt, it does... something. Next time, slightly different something. Hard to debug, hard to fix.
That's the difference. Skill = you can see it, test it, fix it. Agent = black box.
Comment · 10 months ago
Author here. I run an AI company in Tokyo and have been thinking about how Skills (not models) might be the real paradigm shift in the next few years.
The core insight: agents are uncontrollable, Skills are controllable—and that's what makes automation actually deployable.
Happy to discuss. More context on my background: https://x.com/JefferyTatsuya
Comment · 11 months ago
Hey HN, creator here.
Built this after NotebookLM removed their PPT generation feature last week.
Key points: - Free, no login required - BYOK (bring your own Google AI Studio key) for unlimited use - Smart template selection based on content - Built in a week using Claude Code
Tech stack: Gemini API for generation, React frontend.
Examples: https://imgur.com/a/zjZ7oOa
Feedback welcome!
Comment · 11 months ago
Author here. Built this to solve the cost/time problem with Live2D avatars. Traditional approach: 4-5 weeks + $50K-200K. This: <1 day setup, 3-5 sec per clothing change.
Demo: https://avatar.gbase.ai/
Happy to answer technical questions!
I'd spend hours crafting skills for my AI agent - prompts that make it great at reviewing contracts, or workflows that debug code really well. But here's the dilemma: if I share the skill, everyone copies it. If I don't share, it just sits there helping only me.
So I built a marketplace where your agent can use your skills to help others - and you get paid - without ever revealing the skill itself.
Think of it like a chef selling meals, not recipes.
How it works: - Your skills/prompts stay private on your machine - Your agent browses jobs on the platform and applies - It delivers the work, client gets results - You get paid (97% of the job fee)
What I'd love feedback on: 1. Does the "skill privacy" value prop actually make sense to you? 2. What kind of jobs would you post or take? 3. Is this solving a real problem, or am I overthinking it?
Live: https://clawd-work.com Every agent gets $100 free credit to try it out.
It's an experiment - genuinely curious if this model works or if skills should just be open source. Happy to hear both sides.