Help us improve
Share bugs, ideas, or general feedback.
Share bugs, ideas, or general feedback.
Share bugs, ideas, or general feedback.
By aventerica89
Deferred skill loading for Claude Code — compress skills to stubs, load full content on demand. Saves ~88% of per-turn skill token overhead.
npx claudepluginhub aventerica89/skillforgeCompress plugin skills to lightweight stubs, saving ~88% context tokens per turn
Restore compressed skills to their original full content
Show skill compression status and token savings across all plugins
Re-compress skills that lost stubs after plugin updates
Deferred skill loading for Claude Code. Compress skills to stubs, load full content on demand.
Problem: Claude Code injects every skill's description into every system prompt turn. With 60+ skills, that's ~18k tokens of descriptions repeated on every message — context you're paying for but rarely using.
Solution: Replace verbose SKILL.md files with lightweight stubs (~35 tokens each). Full content loads only when a skill is actually invoked.
When a compressed skill is invoked, Claude sees just this:
<!-- SKILLFORGE_COMPRESSED -->
IMPORTANT: Read the full skill before proceeding:
/Users/you/.claude/skills-full/everything-claude-code/api-design.md
That's 3-4 lines instead of the full 524-line skill. Claude reads the stub, follows the redirect, and loads the complete content on demand. One extra Read call per invocation — but since most turns trigger 0-1 skills, the savings are massive.
| Metric | Before | After |
|---|---|---|
| Per-skill description | ~300 tokens | ~35 tokens |
| 60 skills total | ~18,000 tokens/turn | ~2,100 tokens/turn |
| 30-turn session | ~540,000 tokens wasted | ~63,000 tokens |
| Savings | — | ~88% reduction |
/plugin
# Navigate to Marketplaces > jb-claude-plugins > Browse > skillforge
# Or if you have the marketplace: /plugin install skillforge@jb-claude-plugins
# See what's installed and compression status
/skillforge:status
# Compress all plugins
/skillforge:compress
# Compress a specific plugin
/skillforge:compress everything-claude-code
# Restore originals
/skillforge:restore
# Restore a specific plugin
/skillforge:restore everything-claude-code
Restart your Claude Code session after compress/restore for changes to take effect.
SkillForge also provides an MCP server with two tools:
load_skill(name) — Load full content of a compressed skill by namelist_skills() — List all skills in the SkillForge vault~/.claude/skills-full/, outside the plugin cacheClaude Code's skill system injects all skill descriptions into the system prompt on every turn. There's no lazy-loading mechanism for skills (unlike deferred tools, which already support this pattern). We filed anthropics/claude-code#36023 requesting native deferred skill loading. SkillForge is the workaround until that ships.
MIT
Admin access level
Server config contains admin-level keywords
Share bugs, ideas, or general feedback.
Based on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
Skill synchronization and management for Claude Code, Codex, GitHub Copilot, and Cursor. Provides 27 MCP tools for validation, sync, intelligence, and tracing.
Skill to optimize and tune Claude Skills
Create and validate production-grade agent skills with 100-point marketplace grading
Manage Claude Code settings and MCP server configurations with best practices
Create, test, measure, and iteratively improve Claude Code skills with category-aware design, gotchas-driven development, progressive disclosure coaching, and automated description optimization.
Local-only skill quality and security evaluator for Claude Code/OpenClaw. Scores six dimensions, runs local validators, and stores reports and snapshots under .skill-compass/.
Complete 1Password workflow for Claude Code projects — vault setup, secret injection, SAC, CI/CD
Universal Claude Code configuration system with auto-sync across CLI and web. Includes curated commands, agents, skills, and rules for professional development workflows.
Manage local dev OAuth credentials via 1Password. Test any auth provider on any branch without swapping env files.
Deferred skill loading for Claude Code — compress skills to stubs, load full content on demand. Saves ~88% of per-turn skill token overhead.
Zero-copy Obsidian vault over your ~/.claude/ knowledge base via symlinks. One command to make all of Claude's memory, rules, skills, agents, contexts, and plans browsable in Obsidian — no duplication, no sync.
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge.
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge.
Sign in to claim