I spent 2025 talking to Claude Code more than I spoke to humans.
Most of what’s going viral on Twitter today and going mainstream, I was doing three to six months ago.
When I see people using Claude Code to browse the internet, reply to emails, edit their production database using superbase MCP, and edit videos using remotion, I'm not surprised: I've done that.
I Was Living in 2026 While Twitter Was Still in 2025
“Claude Code as DevOps” -- just tell it to use gcloud cli and watch it rip your bill
- “Claude Cowork” -- give Claude root access to your computer and your screen
- Claude Code that lives inside your Postgres
The Internet Isn't for Humans Anymore
Here's a fact that sounds made up: AI agents now generate more internet traffic than humans.
(Source: Imperva Bad Bot Report 2025. Yes, most of it is spam and scrapers. But the trend line is the trend line.)
If you're building a website in 2025 and optimizing for "users," you might be optimizing for the wrong species.
The thing visiting your API endpoint at 3 am isn't a person.
It's Claude running inside someone's automation pipeline, trying to book a flight, summarize a spreadsheet, or do whatever weird task its owner assigned at 11 pm before bed.
Is this important? Like, "reorganizes how you think about software" important? And yet I keep seeing teams design for humans-with-mice when their actual traffic is increasingly agents-with-API-calls.
MCP vs Skills: Top-Down vs Bottom-Up
The reason these things didn't go mainstream earlier is that both the models and prompts were not made for that, and Claude was not able to do it reliably -- you had to prompt it correctly.
If only someone wrote correct prompts for each use case, and shared them in an easy-to-use collection! …Oh wait, that's what skills are, just shareable prompts.
You know the biggest reason that MCP didn't take off as much as it was hyped?
Because MCP was a protocol that was invented top-down and tried to fit reality into it.
Skills, on the contrary, have kind of emerged on Anthropic observing MCPs and how to rip it out, and seeing that Claude performs best when given access to code pieces and small descriptions of what the user is trying to do, with more granular, detailed explanations of how it's usually done.
Anthropic invented MCP to connect multiple services, to unify the interface, and connect services to Claude.
It got a lot of hype, but it was unusable because the moment you combine more than two of them into your AI client, it's eating context like crazy.
Even Manus, the biggest bet of whom were MCP connectors, ended up using programmatic MCP tool calls.
On the contrary, skills were the thing that people were already doing. Usually, they were copying the prompts around and sharing them on Twitter. Personally, I was giving Claude little pieces of code, little tools, and I was asking him to run them in a certain sequence.
For example, here's me in September using Claude to manage my GCP infra.
Three examples
Frontend design skill
How is it possible that you could write one page of prompt and make AI three times better at design? Front-end design skill from Claude one-shots great UIs. While lovable at 100 million ARR still is still not able to produce anything that looks presentable
My guess is that Claude's team worked closely with Opus 4.5, how it thinks, how it writes code. It knows how to make it work hard and burn through tokens. During the front-end design work, Claude Opus brainstorms a lot. It thinks through the content of what you're giving it. It tries to understand how to present it to the user, what would be the matching color palette, what would be the good fonts, it even comes up with icons to use.
All that turns a website into a cohesive piece that clearly communicates the idea
Video editing
Remotion skills launch made 13M views and one-shot people’s brains. You want to know the funniest thing? Everybody was already using it for this
Check out Kate and me winning Best use of Convex with “Remotion skills” back in October 2025 -- we ran Claude Code in e2b sandbox to look through tons of footage and cut it into a nice Reel (github project is here)
Here’s my contribution to Videogen SDK (github link) – it’s a collection of videogeneration APIs and a DSL to put them together. It's open-source.
It might be a good idea for humans, but it’s a fucking catnip for the Claude Code. The guy LOVES this stuff. He’s even more excited than usual.
RALPH WIGGUM
The most useful skill of December 2025 is a one-liner.
“When you finish, observe the state, decide what to do next, and keep working on the next task”.
Claude is famously lazy, and this trick pushes it further into action. This lets me kick off the loop and work on something else, knowing that it will keep working without rest.
What’s next for Agents?
You might say: wait, wasn’t everyone saying that prompt engineering is dead? And now we have even longer prompts?
You need to go up one level. Skills are not prompts. With thinking models, AI prompts itself. The skills are the prompts on how to write THAT self-prompt.
One fun thing I have been doing lately is asking Opus to research me. Instead of writing the prompt, I make it ask questions and pull out from me what I want it to do. I don’t let it start until it has the full picture, though. I call it “reverse agency skill,” and it’s great for when you’re tired and don’t want to make decisions.
My advice: burn tokens. No, really. Just burn as many tokens as possible. You don’t have to burn dollars; just max out your existing subscription at least once.
The interesting stuff happens when you push further. Give it your entire codebase, feed it the full context you'd give a human colleague. Make it read 50 pages of documentation before asking your question. This is how you develop intuitions. Not by reading blog posts (including this one), but by burning tokens until you understand the shape of the thing.
Predictions for 2026 (epistemic status: throwing darts):
- "Agent-first" becomes a real design pattern, not just a buzzword
- Skills proliferate; Vercel were first
- MCP gets quietly deprecated or dramatically rebooted
- At least one major company ships a product where >50% of "users" are AI agents
- I'll look back on this post and be wrong about at least two of these
Will be fun!