跪拜 Guibai
← All articles
AIGC · Claude · AI Programming

Anthropic's Most-Used Internal Skill Is a 321-Byte Prompt

By 王若风 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A 321-byte Markdown file that installs like an npm package shows that the unit of AI productization is shifting from code to precise constraint design. The skill's viral reception also exposes how badly current models default to verbosity—half its value is compensating for the model's own output tendencies.

Summary

ELI5 generates a self-contained HTML artifact that explains any topic—from DNS to a legacy code module—using big pictures and few words, as if to someone who knows nothing. The skill's source code is a 321-byte Markdown file whose core instruction is one sentence: "Explain like I'm someone who knows nothing about this topic, using a HTML artifact with big pictures and few words."

Its effectiveness comes from three constraints packed into that sentence: an absolute beginner audience anchor, a visual-first output format that prevents AI from defaulting to verbose essays, and an HTML carrier that makes diagrams actually renderable. A former Stripe engineer noted the same concept was their most-used internal LLM tool in 2023, though without HTML artifact support at the time.

Commentary on the skill's popularity also surfaced a blunt critique: the model's default verbosity is so bad that a 26-word prompt to force simplicity became a viral product. The skill was committed to the community repository just 40 minutes before the announcement tweet, after five rapid revisions to the prompt wording.

Takeaways
ELI5's entire logic is a 26-word prompt: "Explain like I'm someone who knows nothing about this topic, using a HTML artifact with big pictures and few words."
The prompt anchors the audience as a complete beginner, preventing the model from assuming domain knowledge and producing jargon-dense explanations.
Requiring "big pictures and few words" forces a visual-first output, which avoids the cognitive load of AI's default multi-paragraph essay style.
Specifying an HTML artifact as the output format makes diagrams actually renderable; Markdown alone cannot produce the required visuals.
The skill was committed to the community repository 40 minutes before the announcement tweet, after five rapid revisions to the prompt wording.
Installation requires two CLI commands that add a community plugin marketplace and then install the skill like a package.
A former Stripe employee confirmed the same ELI5/ELI12/ELIPhD concept was their most-used internal LLM tool in 2023, predating Claude's artifact capability.
Highly-upvoted criticism argues the skill's popularity is evidence that the base model's output is too verbose, even when explicitly told to be concise.
Conclusions

The skill's value is split: half comes from well-designed constraints, half from compensating for a model that defaults to overwrought prose. The most-liked comment—"not more intelligence, more intellegible"—frames this as a product debt the model itself carries.

A 321-byte Markdown file installable via a CLI marketplace mirrors the npm/Homebrew packaging model, but the package contents are instructions for AI behavior rather than executable code. This suggests a new product category where the asset is a tested prompt, not a software library.

The five pre-release revisions to a 26-word sentence indicate that prompt engineering at this level is editing, not architecture. The work is removing words until only the constraints that actually prevent a specific failure mode remain.

Concepts & terms
Claude Code plugin marketplace
A community repository maintained by Anthropic where Claude Code skills can be published and installed via CLI commands, analogous to npm or Homebrew but for AI behavior instructions packaged as Markdown files.
HTML artifact
A self-contained HTML file generated by Claude that renders in the browser as a standalone page, supporting layout, SVG graphics, and styling—unlike Markdown, which cannot produce diagrams.
SKILL.md
The single Markdown file that contains all logic for a Claude Code skill, consisting of YAML frontmatter for metadata and a plain-English instruction block that tells the model how to behave.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗