eli5 — a plugin with zero lines of code

Throw it something you don’t know and it explains it again for someone who knows nothing of the field. The name eli5 is short for explain like I’m 5.

It is not a program you install on a computer. Think of it as a note left for the AI assistant I already use, saying “answer like this from now on.” The note contains one sentence, and that one sentence changes the eye level of every answer.

The difference between a hard answer and an easy one is not knowledge — it is the starting point

The same question has two answers. One is written in the language people who already know use; the other is written so that someone who knows nothing can follow from the first line. Both are correct. What differs is where they start.

• The hard answer assumes you know the terms. One unknown word and everything after it is inaudible.
• The easy answer unpacks a term before using it, leans on things you already know, and moves one step at a time.

All this tool does is move that starting point. It does not add new knowledge — it makes the same knowledge get said again from where a newcomer stands.

What installs is not a program but three sheets of paper, weighing about one text message

Opened up: three files, 865 bytes in total. That is hard to picture — it is about the size of a short text message. A single photograph is typically thousands of times bigger.

And of the three, not one line of executable code. All three are paper with words on them.

What is in it	What it does
A note of instruction	Explain as if to someone who knows nothing — effectively the entire product
A name tag	Who made it and which version
One line of usage	A single example

Normally “installing a tool” implies something new starts running. What increased here is one sentence of instruction that the AI reads.

The trigger is decided by prose, not a person

This is the most instructive part of the plugin.

Normally I have to call it by hand — type /eli5 what is DNS. But the note carries a second condition. “When a person asks for a simple explanation, answer this way even if they do not call you by name.”

So the switch is a sentence, not a button. Say explain that simply and the AI reads the condition and answers in this mode on its own.

And yet in three weeks I used it exactly once

This is the real content of the piece.

Three weeks after installing it, I swept my entire working history and counted.

• The tool’s name appears in 944 records.
• The number of times I actually called it: once.

The gap between 944 and 1 exists because the two count different things. 944 is not usage — it is the mark left by the blurb on that note being loaded automatically every time I start the AI. It loads whether I use it or not.

And that one invocation was let’s see what this produces. I was not using the tool, I was looking at it.

	Over three weeks
Cost paid	944 times (loaded automatically, every session)
Benefit collected	0 times (never actually used)

The tool is not at fault — the thought of calling it never arrived

My first reaction was I installed something useless. On a second look, no. The thing was fine. In the moments when an explanation was needed, its name simply never surfaced.

There are three boxes between owning something and using it.

1. Did I install it
2. Did the thought of calling it arrive
3. Did I switch it on

I did 1 and never did 3, and the blockage was at 2. And 2 cannot be solved from the tool’s side however well it is built — it happens inside a person’s head.

So I kept it and changed how it gets called

Two options: delete it, or make it fire without being called.

Deleting was not the answer, because the cost was already being paid — that one line loads every time until it is deleted. Not using it means paying the cost and collecting nothing.

So I went the other way and added a line to my standing rules — when asked to explain something, answer in this mode without being called by name. The switch was taken out of my memory and bolted onto the rules.

What I still do not know

Left here honestly.

• The effect of the change is unmeasured. I changed the rule today, so there are zero days of comparison. A fair comparison needs the same 21 days that came before.
• “Once” is a floor, not an exact figure. Cases where I asked for the same thing in a sentence without naming the tool are not caught by this method.
• I did not judge the quality of the output. This piece is not a record about whether the tool is good — it is a record about the fact that I did not use it.

The detailed record starts here

“Explain it to me like I’m five” — that one instruction, turned into a tool. The name eli5 is short for explain like I’m 5. Throw it a topic you know nothing about — “what is DNS”, “how does this pipeline run” — and it explains it again for someone who knows nothing of the field: unpacking each term before using it, leaning on things you already know.

Delivering that explanation as a single HTML page with big visuals and very few words, rather than as a block of prose, is the form this tool picked. Being easy is the goal; the picture is the means — invert the two and you get difficult material drawn as a diagram. Installing it is one plugin added to Claude Code.

So I opened up what actually gets installed. Three files, 865 bytes, not one line of executable code. What stood out after that was not the thing itself but my own record of using it — in 21 days, its name rode along in 944 of my session logs and it actually ran once.

What it actually is

What gets installed is one skill, and inside it is one sentence. The package holds this and nothing more.

File	Size	Contents
skills/eli5/SKILL.md	331 bytes	name and blurb, a heading, one sentence of instruction, Topic: $ARGUMENTS
.claude-plugin/plugin.json	328 bytes	name, version 1.0.0, author, MIT, five keywords
README.md	206 bytes	one example (/eli5 how does DNS work) and what comes out

The instruction reads: explain this to someone who knows nothing about it, as an HTML page with big pictures and very few words. That is the tool. It ships from anthropics/claude-plugins-community, listed under category: learning.

The wiring is a blurb, not code

Even the trigger is prose. The skill’s description states two conditions — when the user types /eli5 <topic>, or asks for a dead-simple picture explainer. The first is a person calling it by hand. The second says: fire even when nobody calls you.

That is the most useful thing in the package. It is the exact inverse of the teardown where a rule lived only in a README and had zero runtime effect. There, a well-written rule sat where nothing reads it and died. Here, one sentence sits where it does get read and lives. What transfers is not the quality of the sentence but its location.

What the teardown found — 944 mentions, one run

I swept the full session history (2,293 .jsonl files). The string /eli5 appears in 944 of them. That is not evidence of use — it is the mark left by the skill blurb being loaded into every session automatically. Real invocations are recorded differently (a <command-name> marker). Counting those: 1.

And that one run said “show me what this produces.” I was looking at the tool, not using it.

The method matters here. 944 is a ceiling, not usage. The run count of 1 is a floor — asking for the same thing in plain words, without the slash, does not get caught this way. The two numbers are not measuring the same thing, which is why reading it as “used 944 times” gets it exactly backwards.

Held against my own setup — so today I made it standing

The verdict is adopt. Not because the tool is good. The diagnosis came first: the cost of calling it was higher than the benefit.

• One run in 21 days. Not because the thing is weak, but because its name never surfaced at the moment an explanation was needed.
• So I changed how it gets called — asking for an explanation now triggers this style without anyone typing the slash. One line in my standing rules.
• The cost was already sunk. That blurb loads every session anyway. Leave it unused and you pay the cost and get nothing back.

Across the twenty teardowns in this section the recurring verdict has been “I already had it.” This time the diagnosis landed on something I installed myself, 21 days ago. Not someone else’s warehouse. Mine.

What I did not do

• The standing version has zero days of data. I changed it today, so there is nothing to compare against. A fair comparison needs another 21 days.
• I could not separate plain-language calls from slash calls. That is why the run count of 1 is a floor.
• I did not judge the output quality. I saw one page come out; I never built the same explanation both ways and compared them. This teardown is a record of not using something, not a verdict on its worth.