Prompt Libraries Are Dead. Long Live Skills.

Copy-paste prompt collections were the MP3 blogs of AI: scrappy, generous, and structurally doomed. Around 2023 everyone built one. A Notion page with 400 "power prompts." A GitHub repo of awesome-chatgpt-prompts with 100k stars. A $47 Gumroad PDF of "prompts that made me $10k/month." All of them share the same fate as the music blogs of 2006, and for the same reason: the distribution model can't survive contact with a better one.

The better one exists now. It's called a skill, it lives in a Claude Code plugin, and it does everything a prompt library does plus the three things a prompt library structurally cannot: version, attribute, and update itself under you.

What a prompt library actually is

Strip the branding and a prompt library is a pile of untyped strings. You find one, you copy it, you paste it into a chat window, and from that moment it's yours in the worst sense: frozen at the version you copied, severed from its author, and invisible to the tool that's supposed to use it.

That last part is the killer. The model doesn't know your prompt library exists. You are the retrieval system. Every time you need the "senior code reviewer" prompt you have to remember you saved it, find the Notion page, and paste it in. The library's recall rate is your recall rate. Studies of personal knowledge bases put re-retrieval of saved snippets somewhere between "rarely" and "never," and anyone with a bookmarks folder already knows this in their bones.

What a skill is, mechanically

A Claude Code skill is a directory with a SKILL.md at its root. The file opens with YAML frontmatter, two fields matter:

---
name: code-reviewer
description: Invoke when the user asks for a code review,
  mentions "review this diff", or opens a PR discussion.
---

Below the frontmatter is the actual instruction body, and next to it an optional references/ directory for supporting material the skill can pull in. The crucial design decision is that skills load on demand: Claude reads the descriptions, decides when a skill is relevant, and loads the body only then. Your "prompt" stopped being a string you paste and became a capability the agent discovers.

Skills are one of four plugin surfaces in Claude Code, alongside slash commands, hooks, and MCP servers. A plugin bundles them, and a marketplace (a marketplace.json served from a git-backed source, plain HTTP paths silently fail, learn that one the cheap way) makes the whole thing installable:

/plugin marketplace add https://installs.me/lautaro
/plugin install lautaro@lautaro-installs

Two commands. Verified on Claude Code CLI 2.1.x. Compare that to "scroll my Notion page, find section 14, copy the third code block."

MP3 blogs vs. streaming, feature by feature

	Prompt library	Skill in a plugin
Distribution	Copy-paste	/plugin install
Updates	You never see them	Pulled from the marketplace source
Attribution	Lost at first paste	Name is in the frontmatter, source is in the URL
Invocation	You remember to paste	Model triggers on the description
Composition	One string at a time	Skills + commands + hooks + MCP servers

The MP3 blog analogy is precise. MP3 blogs had the content. What killed them wasn't content quality, it was that streaming made the catalog live: versioned, licensed, updated, searchable from inside the player. Spotify didn't beat music blogs by having better songs. It beat them by making the songs installable.

Why the library model lost

Three structural failures, none fixable with a better Notion template.

No versioning. When the author of your favorite prompt learns something and rewrites it, you keep running the old one forever. There is no diff, no changelog, no upgrade path. A skill served from a git-backed marketplace source updates when the author pushes. The people who installed it get the improvement without knowing to look for it.

No attribution. The second a prompt gets pasted into a thread, the author's name falls off. This is why prompt libraries never produced durable reputations; the work was popular and the person was invisible. A plugin carries its author in the install command itself. lautaro@lautaro-installs is a byline that survives distribution.

No context. A prompt is one string pretending to be a person. Real expertise is a system: how someone writes, what they'd advise, which frameworks they reach for, what they'd refuse to do. That doesn't fit in a paste buffer. It fits fine in a plugin: a persona skill for voice, another for decision frameworks, references/ full of the actual source material, all loading only when relevant so it doesn't torch your context window.

The persona is the unit, not the prompt

This is the part the prompt-library era got backwards. People never actually wanted "the 10 best prompts for founders." They wanted the founder. The prompt was a lossy compression of a person's judgment, and we tolerated the loss because there was no way to ship the person.

Now there is. installs.me takes a creator's files, calendar, and calls and synthesizes them into a persona plugin: skills that write in their voice, advise with their frameworks, and cite their actual history. The demo persona is a real one, Lautaro Schiaffino, Argentine serial founder, sold Sirena to Zenvia for $30M, now running Darwin AI. His plugin doesn't hand you a "think like a founder" prompt. It installs the specific founder, and when he ships an update, everyone who installed him gets it.

MP3 blogs were a beautiful, doomed decade. Save your prompt library as a keepsake. Then package it.

Install a person

installs.me turns your files, calendar and calls into a Claude Code plugin that thinks like you. Anyone installs it with two commands:

/plugin marketplace add https://installs.me/lautaro
/plugin install lautaro@lautaro-installs