collection

collection

A Claude Code plugin built by One Studio Co.

Origin

We built a multi-agent system in Claude Code — a command center with 10+ specialized agents, each with their own memory, tools, and domain. It worked well. Then we noticed something: spawning a simple agent just to answer a one-line question cost ~32,000 input tokens. Every single time.

We traced it. The agent wasn't doing anything special — it just loaded CLAUDE.md (~9,600 tokens) + MEMORY.md (~2,800 tokens) + its own agent file (~4,900 tokens) + its memory index (~2,000 tokens) before saying a word. Files that had grown organically over months, full of detail that was only occasionally needed.

The fix was obvious in hindsight: the same thing Claude Code already does to its own conversation history — compress. Keep the essential summary always in context, move the raw detail somewhere agents can Read on demand.

We ran a full audit and compression pass across the whole system:

What	Before	After	Reduction
CLAUDE.md	38,519 bytes	5,630 bytes	-85%
Average agent file	~10,000 bytes	~4,800 bytes	-52%
Average memory index	~5,000 bytes	~1,700 bytes	-66%
Heaviest agent spawn	~15,700 tokens	~4,700 tokens	-70%

One agent was running on Opus when Sonnet was sufficient — fixing that alone cut its cost 5× per session.

We turned the process into /slim so any Claude Code agent system can run the same pass.

---

What are Claude Code skills?

Skills are reusable prompts that extend Claude Code with domain-specific workflows. Drop a skill file into .claude/commands/ and invoke it with /skill-name in any Claude Code session.

Skills

/slim — Context Budget Auditor

Measures and compresses the token cost of your Claude Code agent system.

Every agent spawn loads CLAUDE.md + MEMORY.md + the agent file + its memory index — all verbatim. As projects grow, these files accumulate detail that's only needed occasionally. slim audits all always-loaded files, proposes a two-tier split (fast-tier always loaded, slow-tier read on demand), and compresses agent files and memory indexes — without losing any content.

Run it when:

• Spawning an agent costs more tokens than expected
• CLAUDE.md has grown beyond ~5KB
• Agent files have accumulated protocol details that duplicate CLAUDE.md
• You want a monthly context hygiene pass

What it does (6 phases):

1. Audit — measures every always-loaded file, prints a token cost table sorted by impact
2. CLAUDE.md split — extracts reference material to docs/claude/*.md, keeps only the fast-tier in CLAUDE.md
3. MEMORY.md compression — collapses verbose entries to one-liners, removes dead links
4. Agent file compression — removes duplication with CLAUDE.md, cuts prose rationale, keeps domain-specific rules
5. Memory index compression — each entry to one line, under 80 lines per index
6. Model right-sizing — audits model: frontmatter per agent, recommends Haiku/Sonnet/Opus based on task type, documents a dynamic escalation pattern for edge cases

Interactive — proposes changes, waits for approval before writing each phase.

Installation

This plugin is distributed via the One Studio marketplace on GitHub.

1. Add the marketplace

From within Claude Code:

/plugin marketplace add onestudio-co/claude-code

Or from the CLI:

claude plugin marketplace add onestudio-co/claude-code

2. Install the plugin

From within Claude Code:

/plugin install collection@one-studio

3. Restart the session

After installing, restart your Claude Code session (/exit then relaunch) to load the new plugins.

Note: The built-in /reload-plugins command should reload newly added plugins without a restart, but it currently does not. This is a known bug. Once the Anthropic team fixes it, you can use /reload-plugins instead of restarting.

4. Use the skills

/slim

Update

To pull the latest version of the plugin, update the marketplace:

From within Claude Code:

/plugin marketplace update one-studio

Or from the CLI:

claude plugin marketplace update one-studio

License

MIT