What I took apartpublic

eli5 — a plugin with zero lines of code

One instruction turned into a plugin — explain it to me like I am five. Opened up, it is three files, 865 bytes, zero lines of executable code — and in 21 days I called it exactly once.

865 bytes · 1 rungithub.com/anthropics/claude-plugins-community ↗

“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 using it looks like — ask in one line, hear it back at a beginner's level what I ask "what is DNS" "how does this pipeline run" one line is enough → eli5 "explain to someone who knows nothing" one sentence is the whole tool → what comes out — one plain explanation big visuals, few words ⇒ The whole tool is the level — an expert's baseline swapped for a beginner's The picture is how it gets there — terms get unpacked before they get used
What this tool sells is not accuracy but a lower doorstep — letting someone who knows nothing get past the first page.

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.

21 days — times the name was carried vs. times it ran session logs carrying the name 944 recorded invocations 1 944 is not usage — it is the blurb being loaded automatically every session That single run asked what it produces. Looking, not using. ⇒ The standing cost was paid 944 times; the benefit was collected zero times Not the tool's fault. The moment of remembering it never arrived.
Between installed and used there is one more box: did it occur to me to call 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.