15 years of writing code, and I’ll tell you what just changed for me.
It’s not the model. The models are good. They’ve been good.
It’s Skills.
Garry Tan, CEO of Y Combinator, shipped 600,000 lines of production code in the last 60 days — part-time, while running YC.
Austin Lau ran Anthropic’s entire growth marketing operation solo for ten months, and he’d never opened a terminal before he started.
Peter Steinberger built OpenClaw — 250,000 GitHub stars in three months, fastest-growing open-source project in history.
Different jobs. Same lever. Skills.
I spent a week tearing through every Claude Skill I could find — public ones, private ones, weird ones nobody’s posting about. These are the 5 I now refuse to work without.
1. Content Ops — Eric Siu
Most people ask Claude for ten headlines, skim them, pick the one that feels least bad, ship it. That’s not writing. You’re letting the model pick.
Eric’s Skill flips it. You generate fifty. An expert panel — copywriters, conversion specialists, landing page strategists — scores every one. Anything below 90 dies. Survivors get re-mutated. Three rounds in, you’re not looking at your headline. You’re looking at one that beat 49 versions of itself.
This is Karpathy-style evolution applied to copy. Real statistical scoring. Bootstrap confidence intervals, Mann-Whitney U tests. Not vibes.
If you ship words to humans, this is the first install.
→ github.com/ericosiu/ai-marketing-skills
Inside the community
I’m building a set of AI skills and tools for people building real businesses
Interactive Site Builder, YouTube Script Skill, and Transcript Grabber are the first releases — plus the rest of the toolkit I use to build and run my agency. First 200 members get in at the early bird price.
2. TradingAgents — Tauric Research
64,000 stars on GitHub, and the architecture is the gift.
Give it a ticker. A fundamentals analyst pulls financials. A sentiment expert reads the news. A technical analyst looks at the chart. They argue. A research manager synthesizes. A trader proposes a position. A risk team — multiple agents with different risk appetites — pressure-tests it. A portfolio manager makes the final call.
It’s a multi-agent simulation of an actual trading desk. But the part I keep stealing is the pattern. Multiple expert lenses on the same decision, each one an agent, all of them debating. Swap the agents and the same loop runs on hiring decisions, investment memos, product bets.
(Disclaimer: it’s a research framework. Not financial advice. Don’t YOLO your savings.)
→ github.com/TauricResearch/TradingAgents
3. Twitter Algorithm Optimizer — Composio
Twitter open-sourced its recommendation algorithm a while back. Most people read the headline and moved on. Composio actually read the source code.
Then they built a Skill on top of it.
Drop in a tweet. The Skill checks it against the four real models that decide whether your tweet goes anywhere — Real-graph, SimClusters, TwHIN, Tweepcred — tells you which mechanism it’s failing on, and rewrites it.
A flat opinion like “remote work is better than office work” comes back as a hot take with a clear position, a debate hook, a question, and a community signal. Same idea. Different engagement profile.
This is what algorithm-aware writing looks like.
→ github.com/ComposioHQ/awesome-claude-skills
4. gstack — Garry Tan
Garry Tan, CEO of Y Combinator, shipped 600,000 lines of production code in 60 days. Part-time. While running YC.
The thing he uses to do it is open-source, MIT-licensed, installs in 30 seconds. It’s called gstack.
The move is brilliant. He didn’t build one giant agent. He built a team. /office-hours is a CEO agent that pushes back on your framing before you write a line of code. /plan-eng-review is an engineering manager that locks the architecture and writes the test matrix. /qa opens an actual Chromium browser, clicks through your app like a user, finds bugs, fixes them with atomic commits, and writes a regression test for every fix.
23 commands. Then the cheat code: he runs ten to fifteen of these loops in parallel.
This is what a one-person YC startup looks like in 2026.
5. Remotion — remotion-dev
Try this. Open Claude. Ask it to write a Remotion component cold. Run it. Watch it crash at frame 47.
Cold Claude gets it almost right and then breaks in three subtle ways. CSS transitions that don’t render. Tailwind animation classes that fail silently. Frame counts that miscount. Looks fine in the browser, produces black video.
The Remotion team shipped an official Skill with 30+ rule files. Text animations, transitions, font loading, frame measurement. With it installed, Claude stops guessing. It uses useCurrentFrame and interpolate properly. It knows when to use spring timing versus linear. It loads assets with staticFile.
If you build any kind of programmatic video, this is the difference between Claude shipping a working component and Claude shipping something that crashes the renderer.
→ github.com/remotion-dev/skills
The pattern
Five Skills. Marketing, trading, social, engineering, video. Different jobs, same shape.
The great Skills aren’t prompts. They’re workflows with an opinion about how the work should be done.
— Sean