Discovering, curating and stacking skills

Skills are discovered in two public libraries maintained by Anthropic, and several skills declared for distinct needs can load together on the same task, but a poorly chosen description never triggers, whether alone or stacked.

An installed skill is useless if its description never overlaps with the vocabulary of the requests aimed at it: looking for an existing skill before writing a new one first avoids this risk, since a skill already widely used generally has a description already tried and tested against varied requests. Two public libraries, maintained by Anthropic, cover the bulk of this search.

Two libraries, two levels of curation

The first, claude-plugins-official, is curated directly by Anthropic; it registers automatically on the tool's first interactive launch, or can be added by hand. The second, claude-community, gathers third-party submissions reviewed before publication, and is added explicitly with a marketplace command. Neither publishes an official figure for the number of skills it holds at any given moment, that figure changes too fast for a value printed in this course to stay true for long. Curation, in both cases, does not replace a careful reading of the description before installation: a widely adopted skill can still poorly cover the precise need it is meant for.

# Exemple illustratif, ajout d'une bibliothèque communautaire
/plugin marketplace add anthropics/claude-plugins-community
# puis, une fois la bibliothèque enregistrée, parcourir son catalogue
/plugin search changelog

Stacking without letting the context load for nothing

Nothing prevents several skills, each declared for a distinct need, from loading together on the same task: each is evaluated independently against the text of the request. What limits stacking, then, is not an imposed ceiling, it is the real cost, every installed skill adds text to the context of every session, even when it does not trigger. Curating without mercy means removing a skill as soon as its description no longer matches a genuine recurring need, rather than letting it silently accumulate weight. In the event of a name shared between two sources, priority goes first to a company skill, then to a personal skill, then to a project skill, which settles unambiguously which file applies.

The precondition for all this searching remains the same as for writing: a precise description and a correctly formed frontmatter, exactly what creating a skill that is also a command describes. A rich library never compensates for a poorly written description, it only multiplies the number of missed opportunities.

Figure 1

The two public skill libraries

Skill library	Curation	Registration	Origin of submissions
claude-plugins-official	Curated directly by Anthropic	Automatic on first interactive launch, or manual	Content selected by Anthropic
claude-community	Third-party submissions reviewed before publication	Added explicitly with a marketplace command	Content proposed by third parties

Both libraries cover the same discovery function, but with different filtering before publication.

Figure 2

Installing, keeping or removing a skill

Check whether an existing skill already covers the need before writing a new one.

Install

Add the skill from a public library or write it yourself if nothing matches.

Test

Check on a real case that the description triggers the skill at the right moment.

Remove

Uninstall the skill as soon as its description no longer matches a genuine recurring need.

The cycle starts from a genuine recurring need and closes with a removal as soon as the skill is no longer useful for that need.

Calibrate it yourself

A developer adds the community skill library to his Claude Code session, installs a skill intended to generate release notes, and uses it on three successive requests concerning three different repositories.

Write, in one sentence, what this situation establishes, and in one sentence what it does not establish.

Your answer, before you see the key

What to remember

Two public libraries maintained by Anthropic cover skill discovery, one curated directly, the other open to reviewed third-party submissions.
Several skills can load together on the same task, each evaluated independently against the text of the request.
Every installed skill adds text to the context of every session even when it does not trigger, which justifies removing those that no longer match a real need.
In the event of a name shared across several sources, priority goes first to a company skill, then to a personal skill, then to a project skill.

Do this now

Open the skill libraries panel in a Claude Code session, add the community library if it is not already there, and browse its catalogue to spot a skill that covers a need you genuinely face this week.

Check the source

Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.