Anthropic's One-Sentence ELI5 Skill Redefines AI Output as Interactive HTML

Hello, I'm Wai Wai.

Over the weekend, I saw a tweet on X with a staggering 1.66 million views:

The tweet's content was very simple; it just mentioned a skill called eli5.

Then it explained the skill's function:

Explain like I'm someone who knows nothing about this topic, using a HTML artifact with big pictures and few words.

This tweet spread quickly.

Why?

Because of the tweet's first sentence: a skill people at Anthropic have been using a lot recently: ELI5.

This mentioned that people at Anthropic have been frequently using this skill recently.

And the guy who posted the tweet, Thariq, is one of the founding team members of Anthropic, and he has a large following on X:

So, the skill he mentioned quickly spread.

After seeing it, I also gave it a try.

How was the experience?

It was subtle.

Skill Content

The installation method for Claude was provided below the tweet:

claude plugin marketplace add anthropics/claude-plugins-community

claude plugin install eli5@claude-community

If your tool isn't Claude, you can also find this skill on GitHub:

https://github.com/anthropics/claude-plugins-community/blob/main/eli5/skills/eli5/SKILL.md

Honestly, when I saw the content of this skill, I felt like I had been "played."

Because its complete content is just one sentence:

Explain like I'm someone who knows nothing about this topic, using a HTML artifact with big pictures and few words.

This sentence is exactly the one I saw in the tweet earlier:

I initially thought this was just a description of the skill.

I never imagined it was the entirety of the skill.

How the hell can this be considered a skill?

The last time I had this feeling was when I saw the "grill-me" skill:

That skill was only three sentences, but after trying it a few times, I thought those three sentences were really awesome.

So, although "eli5" gave me the same feeling in terms of skill content this time, I didn't dare to judge it casually before actually trying it.

I just kept a small question mark in my heart.

Also, look at the first sentence of this skill's description: Explain a topic like I'm a 5 year old.

Please explain a topic in a way suitable for a 5-year-old child.

I have enough reason to reasonably guess that the name of the skill "eli5" is an abbreviation derived from this sentence.

Trying it out

After installing the skill, the first question I asked it to explain was:

The working mechanism of "text watermarking."

Because I specifically researched this topic some time ago and even wrote an article about it, I roughly know how it works.

I first tried it with a topic I'm familiar with to see if it would talk nonsense:

This is the result it outputted, an HTML page. Let me quickly scroll through it to show you the complete content:

If I took screenshots, the content would be too long, so I'll just capture a few key images:

If you've read my previous article "Eye-opening! Articles written by AI might all carry 'invisible watermarks.' It's still quite fierce.", you'll know that the content in the previous images is quite accurate and hits the key points.

Additionally, in the HTML it generated, it actually made an interactive button for me, which I didn't expect:

Judging from this case, the feeling was pretty good.

But actually, this was the HTML I got after asking the same question a second time.

This was its first answer to this question:

Although the first answer also mentioned the key points, I personally felt there was still a gap compared to the second one.

But asking the same question twice wasn't intentional on my part.

It was because the first time I used it, I was in the wrong directory, so I asked again after switching directories.

That's when I discovered its "gacha" mechanism.

Because the questions I thought of always leaned towards technical directions, I asked my friend Max to think of a question for me.

She indeed came up with a very good question:

The principles and functions of the Dujiangyan irrigation system

The question was a good one, but I felt eli5's answer was very mediocre.

Let me first show you a GIF to give you a feel for the overall content of the HTML it generated:

It did mention key parts like "Fish Mouth, Flying Sand Weir, and Precious Bottle Neck":

But the two diagrams that followed, I couldn't figure out what they were "illustrating" at all:

Inside Claude, it guided me, asking if I wanted to generate another screen about "Dig the channel deep, keep the weir low":

I let it add it, and this was its answer:

Perhaps because I know Dujiangyan relatively well, I wasn't really satisfied with this answer.

I felt its answer couldn't clearly explain the awesomeness of Dujiangyan to someone who knows nothing about it.

Since we're on the topic of Dujiangyan, I'll casually recommend a video that I think explains Dujiangyan clearly enough.

This video by the Institute of Planets on Bilibili; if you're going to "pay respects to the water at Dujiangyan," you can do your homework in advance:

https://www.bilibili.com/video/BV1JK42147uN/

Alright, I've digressed a bit too far.

Let's get back to the eli5 skill.

Finally, I asked it a question very suitable for it:

Demonstrate the "Monty Hall problem."

First, what is the "Monty Hall problem"?

It's actually a counter-intuitive probability game.

You're asked to pick one of three doors, behind one of which is a car, and behind the other two are goats.

Then the host deliberately opens one of the doors you didn't pick that has a goat behind it, and asks if you want to switch doors.

Intuitively, with only two doors left, the probability seems like a 50/50 split.

But what's the actual situation?

eli5 will tell you in a moment.

Also, why did I say this question is very suitable for eli5?

Because its answer is HTML, and you can do interactions in HTML; it could completely simulate the "Monty Hall problem" game:

This is the HTML page it gave. First, let's look at its overall content:

It created an interactive page inside, first asking me to choose a door. I chose door number 2:

Then, it told me the host opened door number 3, revealing a goat.

If I insisted on not switching, I'd get a goat. If I switched, I'd get the car:

Then it "let the data speak," simulating switching vs. not switching 10,000 times each.

Multiple simulations showed that the probability of winning by switching is much higher than not switching:

At the end of the HTML, it also described why the winning rate is higher when switching:

I thought this case was pretty good.

It achieved the goal of explaining the problem to the user through interaction.

Furthermore, here are a few scenarios Thariq mentioned he uses frequently:

If you have similar scenarios at work, you can give it a try.

Very Subtle

As I mentioned earlier, my feeling after trying it was: very subtle.

First, let me complain about the bad points.

Based on the experiences above, it didn't leave a particularly good impression on me.

Maybe my way of using it was wrong, but that's genuinely how I felt.

Its ability to spread widely is largely due to the backing of Company A (Anthropic).

If this was a Skill written by an unknown person, consisting of just one sentence, and the actual experience wasn't that amazing,

I would only leave a comment: What the hell is this thing?

But when this Skill was written by an engineer from Company A, and they also indicated they've been using it a lot recently,

I would think: this skill is profound, I probably haven't grasped its deeper meaning yet; my cultivation is insufficient.

I think the rapid spread of this Skill is mostly attributable to the halo effect of Company A.

Note, I said "mostly."

There's another part, which is because it genuinely has something to it.

That is the point that made me feel good.

It indeed hit upon a very good direction: making AI not just output text, but use HTML to create a truly human-oriented, visual explanation page.

HTML, in the current AI era, is a very "cheap" yet highly suitable medium for displaying content.

There were even some debates online previously about whether HTML should replace md format documents.

And the starting point of that debate was coincidentally also an article published by this guy, Thariq:

For those supporting HTML, their main points were like this:

And for those supporting md, their points were like this:

In the case of eli5, its deliverable is HTML, which I actually think is great, because HTML naturally possesses richer visual expression and interactive capabilities.

This is also a point I learned from the eli5 skill.

When the target user of the output is a human, HTML is indeed a better choice.

Additionally, if we look purely at the Skill file itself, it is indeed the simplest Skill I have ever seen, bar none.

But this experience with this Skill made me re-understand a problem: the value of a Skill might never lie in how many words it has, how long or detailed it is.

A single good instruction is sometimes enough to change what we think of as "AI's way of working."

Previously, when we asked AI to "explain a problem," the default deliverable was a piece of text.

But eli5 overturned this way of working with just one sentence, making it create something.

Inside this something, it can even embed interactive parts based on the question, allowing the user to better understand the problem.

So, I think what's truly worth learning might not be that this Skill is just "one sentence," but the redefinition of AI deliverables behind that one sentence.

Finally, I strongly recommend you try out the eli5 skill yourself.

If you find a better way to use it and get a better experience, feel free to leave your thoughts in the comments section.