By skyforbes 
Okay so a lot of us spent last year obsessing over context windows.
200K tokens! 1M tokens! Google announced “infinite context” which… sure. The assumption was bigger = better, and every AI lab threw engineers at the problem.
Then Anthropic did something weird. Instead of making their context window bigger, they released Skills in mid-October. And honestly? For certain use cases, it’s smarter than what everyone else is doing.
Not revolutionary. Just smarter.
What They Actually Built
A skill is just a folder with a markdown file in it. Sometimes there’s code or templates too, but the core is stupidly simple: a SKILL.md file that tells Claude how to do specific things.
Want Claude to format reports your company’s way? Make a skill. Need code reviews that follow your team’s standards? Another skill. Presentations with your brand guidelines baked in? Skill.
Here’s the part that matters: Claude doesn’t load all this stuff into memory at once. It scans the metadata (like 30–40 tokens per skill), figures out what’s available, and only loads the full instructions when you actually need them.
Compare this to GPTs where you pick one and it loads everything. Or Projects in Claude where all your context sits there eating tokens. Skills just… wait in the background until they’re useful.
More efficient. Less flexible. Pick your poison.
Who’s Actually Using This
Box is using Skills to auto-format files into presentations and spreadsheets. Notion integrated them. Canva’s testing them. Rakuten — which is massive — is apparently running trials.
The numbers they’re sharing:
- 60–70% faster document formatting (when you have strict brand standards)
- Code reviews taking 40% fewer rounds of feedback
- New developers onboarding 2–3x faster
- Compliance rates went from like 60% to basically 100%
One finance team said they cut a day-long process down to an hour. Spreadsheet analysis, anomaly detection, report generation — all automated using their procedures.
Now. Are these cherry-picked success stories? Obviously. Companies that tried Skills and hated them aren’t giving interviews. But the pattern is consistent enough to pay attention to.
Why This Isn’t Just “GPTs But Different”
OpenAI built the GPT Store as a marketplace play. Maximum distribution, maximum creators, maximum network effects. App store model for AI.
Makes sense if you want ecosystem growth. More GPTs → more users → more GPTs. Classic platform dynamics.
Anthropic went the other direction. Skills are a control play.
They work the same across Claude.ai, Claude Code, and the API. There’s approval gates before actions execute. Everything’s versioned. It’s designed for enterprises that cannot tolerate unpredictability.
And look — if you’re in finance or healthcare or legal, you literally cannot deploy AI that makes random decisions. You need defined capabilities, audit trails, version control. Skills give you that.
But if you’re building consumer products or you need maximum flexibility? GPTs make way more sense. Bigger ecosystem, more integrations, faster iteration.
Neither is “better.” They’re optimized for different customers.
What This Actually Changes for PMs
If you’re building products with AI, a few things shift:
You need your standards documented first. Skills can’t help if your workflows live in people’s heads. You have to actually write down how things work before you can encode them.
Measure consistency, not magic. The value prop isn’t “AI did something amazing.” It’s “AI did the right thing reliably.” Track error rates and compliance metrics.
Build modular, not monolithic. Skills stack. One analyzes data, another visualizes it, a third builds the deck. Think components, not one giant AI system.
Version control matters now. You can test Skill changes before deploying them. You can roll back if something breaks. Treat your AI capabilities like code releases.
Some companies are building Skill libraries that capture institutional knowledge. Not using Skills for one-off tasks — building up a collection of “this is how we do things” that the AI can reliably execute.
That’s infrastructure, not features.
The Real Limitations (That Marketing Won’t Tell You)
Skills have actual constraints:
The ecosystem is smaller than GPTs. Fewer integrations, fewer examples, less community support.
They’re optimized for defined tasks. If you need open-ended exploration or creative problem-solving, Skills get in the way more than they help.
Setup cost is real. You have to document and structure your workflows upfront. That takes time.
And if your workflows change constantly? Maintaining Skills becomes another thing to maintain. You’re trading flexibility for consistency.
Plus the big one: you have less control over when Skills activate. Claude decides when to use them based on your request. Sometimes it picks the right one. Sometimes it doesn’t. You can’t always predict behavior.
Does Token Efficiency Even Matter Anymore?
Here’s the question I keep coming back to: if every model gets 10M token context windows, does efficiency still matter?
Maybe. But not for the reasons you’d think.
The problem shifts from “can the model remember this?” to “does the user know what the model knows?”
When you have 50 GPTs installed, which one do you use? When your project has 15 different instructions active, what’s actually affecting the output? More capacity creates new UI problems.
Skills solve discovery by being invisible. Claude picks them automatically based on what you’re doing. Better UX, but you give up explicit control.
Some people will hate that. Some will prefer it. Depends on your use case.
Where This Goes
The market’s splitting:
One segment wants maximum capability. They’ll use GPTs, customize agents, build on open models. Mostly consumer products and startups that need flexibility.
Another segment wants predictable behavior. They’ll use Skills, deploy on their own infrastructure, maintain governance. Mostly regulated industries and big enterprises.
First segment is bigger but price-sensitive. Second segment is smaller but pays more for reliability.
We’ve seen this pattern before. MySQL vs Oracle. Open source vs commercial. Consumer web vs enterprise. The volume play and the margin play need different products.
Anthropic’s betting on margins. OpenAI’s betting on volume. Both can work.
Bottom line: Skills are good for defined, repeatable tasks in environments that need control. They’re not great for exploration, creativity, or situations where flexibility matters more than consistency.
If you’re serving regulated enterprises, Skills might be worth it. If you need maximum flexibility or you’re building consumer products, GPTs make more sense.
It’s not some huge breakthrough. Just a different way to solve the problem of making AI useful for specific organizational workflows.
Some companies need that. Some don’t.
Honestly the most interesting part isn’t the technology — it’s the strategy signal. Anthropic looked at the context window arms race and said “we’re playing a different game.” That tells you more about their roadmap than any feature announcement.
Whether it works long-term? We’ll see. But at least they’re trying something different instead of just making the same thing bigger.