Skip to content

Claude Code

Skill Atlas cuts Claude Code’s skill overhead from 2,544 tokens to 338

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Claude Code versions and changes →
Image: pexels.com

Skill Atlas is a Claude Code plugin that keeps six skills loaded and hides the other 47 behind a search. On that 53-skill collection, the fixed cost of skills fell from 2,544 tokens per session to 338. The figures come from the project’s README on GitHub, published August 14 by the developer danielLublinsky. They describe one setup, and every search is billed separately on top of them.

The 338 covers the six enabled skills’ descriptions plus one line of about 50 tokens at session start that tells Claude the dormant tier exists. The other 47 skills cost nothing until something asks for them.

A tier between on and off

The plugin works around how Claude Code loads skills. The Claude Code docs say a skill’s body loads only when the skill is used. Its name and description are loaded every session, because Claude matches each request against those descriptions to decide which skill to run. The README puts that at about 48 tokens per skill. Past 100 skills, the author writes, that adds up to thousands of tokens a session, “and near-duplicates quietly compete for the model’s attention.”

“Disabling fixes the cost and loses the skill,” the README says. Skill Atlas adds a middle state. Enabled skills behave as before. Searchable skills keep their descriptions out of context but can still be found through a skill-search skill. Disabled skills can’t be found at all. You move plugins between tiers with /skill-atlas:edit-searchable, and you set everything up once with /skill-atlas, which builds the graph, sorts new skills into categories and renders an HTML map. After that, a SessionStart hook keeps it current.

Day to day, you rarely type the search command. Claude calls it before a nontrivial task, names the skill it picked in one line and carries on. Installing takes two commands: claude plugin marketplace add danielLublinsky/Skill_Atlas, then claude plugin install skill-atlas@skill-atlas. The plugin needs only Python 3, has no dependencies, makes no network calls and is licensed under Apache-2.0.

What a search costs

Search never loads the whole catalog. In the README’s example, a task like “resolve this merge conflict” has Claude read _index.md, a 14-line file that lists each category’s name, member count, scope and token cost, for about 1,600 tokens. Claude then picks one category file, called a shard, for about 600 more. In this case it’s version-control.md, which holds the merge-conflict skill. The other 13 shards, about 10,000 tokens together, stay closed. Add the search skill’s own 400 tokens and a lookup comes to about 2,400. If the match is an enabled skill, Claude invokes it normally. If it’s dormant, Claude reads the skill file at its path and follows it as instructions.

That 2,400 is paid each time a search runs, while the dormant descriptions it replaces were billed from session start whether you searched or not. The author is upfront about the trade: “The arbitrage is real when searches are occasional; it is not free, and the numbers above are measured rather than estimated.”

The number of reads is capped. One shard is the norm. A second is opened only when the entry Claude found points to a category it hasn’t read yet, which the README describes as “the catalog pointing, not the model guessing.” Claude never opens a third. Skills appear in more than one category on purpose, so Claude only has to land on a category that contains the right skill. Disabled skills are left out when shards are written, so a plugin you turned off can’t come back through search.

Categories frozen after the first run

Claude categorizes the collection once. On the first run it drafts 8-12 categories named after the kinds of task users bring, and after that the taxonomy is fixed. Later runs only file new skills and recheck those whose descriptions changed. Each assignment stores a hash of the description it was based on, so the plugin can spot stale assignments without re-sorting the whole library. Search results are announced in one line before work continues. In the README’s words, “search is a lookup inside a task, not a deliverable.”

The graph underneath

The other half of the plugin is a graph of your skills, built for debugging them. Discovery starts from the manifests: installed_plugins.json, then each plugin.json, then settings. That keeps marketplace catalogs and stale cached plugin versions out of the count. The obvious shortcut, treating every directory that contains a SKILL.md as a skill, overcounts by about 60% on a real machine, the README says, and build_graph.py --naive-count runs that check on your own setup.

Each skill is a node, and there are two kinds of edges. A references edge links a skill to its own bundled files and turns red when a file is missing, so a mistyped references/foo.md shows up without opening anything. A mentions edge records one skill naming another and is marked dangling when the target is disabled, unregistered or absent. Mention matching is deliberately strict: only names in backticks, skills/<name> paths and unambiguous hyphenated names count, and code fences are stripped first. Looser matching produced 91 edges on a 41-skill collection, almost all of them ordinary English words.

The graph also reports duplicate names and orphaned skills, and it returns an exit code you can use to fail a CI run. If you skip the plugin, make check exits with 1 on any broken reference or dangling mention. atlas.html is a single file with D3 inlined, and it opens from file:// without any network requests. Each node’s fill shows its tier. Duplicate names are listed in the footer with both nodes drawn, and you can pin any skill or bundled file and read its markdown source in a popup without leaving the graph.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.