context-loader
About
This skill explains how SpecWeave leverages Claude's native progressive disclosure (skills) and sub-agent parallelization for efficient context management. It helps developers understand token usage and scaling for large projects without custom caching systems. Use it when asking about context efficiency or how SpecWeave handles large codebases.
Quick Install
Claude Code
Recommended/plugin add https://github.com/majiayu000/claude-skill-registrygit clone https://github.com/majiayu000/claude-skill-registry.git ~/.claude/skills/context-loaderCopy and paste this command in Claude Code to install this skill
Documentation
Context Management in SpecWeave
Overview
SpecWeave achieves efficient context usage through two native Claude Code mechanisms:
- Progressive Disclosure (Skills) - Claude's built-in skill loading system
- Sub-Agent Parallelization - Isolated context windows for parallel work
Important: SpecWeave does NOT use custom context manifests or caching systems. It leverages Claude's native capabilities.
1. Progressive Disclosure (Skills)
How It Works
Claude Code uses a two-level progressive disclosure system for skills:
Level 1: Metadata Only (Always Loaded)
---
name: nextjs
description: NextJS 14+ implementation specialist. Creates App Router projects...
---
What Claude sees initially:
- Only the YAML frontmatter (name + description)
- ~50-100 tokens per skill
- All skills' metadata is visible
- Claude can decide which skills are relevant
Level 2: Full Skill Content (Loaded On-Demand)
# NextJS Skill
[Full documentation, examples, best practices...]
[Could be 5,000+ tokens]
What Claude loads:
- Full SKILL.md content only if skill is relevant to current task
- Prevents loading 35+ skills (175,000+ tokens) when you only need 2-3
- This is the actual mechanism that saves tokens
Example Workflow
User: "Create a Next.js authentication page"
↓
Claude reviews skill metadata (35 skills × 75 tokens = 2,625 tokens)
↓
Claude determines relevant skills:
- nextjs (matches "Next.js")
- frontend (matches "page")
- (NOT loading: python-backend, devops, hetzner-provisioner, etc.)
↓
Claude loads ONLY relevant skills:
- nextjs: 5,234 tokens
- frontend: 3,891 tokens
↓
Total loaded: 9,125 tokens (vs 175,000+ if loading all skills)
Token reduction: ~95%
References
"Skills work through progressive disclosure—Claude determines which Skills are relevant and loads the information it needs to complete that task, helping to prevent context window overload."
2. Sub-Agent Parallelization
How It Works
Sub-agents in Claude Code have isolated context windows:
Main conversation (100K tokens used)
↓
Launches 3 sub-agents in parallel
↓
├─ Sub-agent 1: Fresh context (0K tokens used)
├─ Sub-agent 2: Fresh context (0K tokens used)
└─ Sub-agent 3: Fresh context (0K tokens used)
Benefits:
-
Context Isolation
- Each sub-agent starts with empty context
- Doesn't inherit main conversation's 100K tokens
- Can load its own relevant skills
-
Parallelization
- Multiple agents work simultaneously
- Each with own context budget
- Results merged back to main conversation
-
Token Multiplication
- Main: 200K token limit
- Sub-agent 1: 200K token limit
- Sub-agent 2: 200K token limit
- Effective capacity: 600K+ tokens across parallel work
Example Workflow
User: "Build a full-stack Next.js app with auth, payments, and admin"
↓
Main conversation launches 3 sub-agents in parallel:
↓
├─ Sub-agent 1 (Frontend)
│ - Loads: nextjs, frontend skills
│ - Context: 12K tokens
│ - Implements: Auth UI, payment forms
│
├─ Sub-agent 2 (Backend)
│ - Loads: nodejs-backend, security skills
│ - Context: 15K tokens
│ - Implements: API routes, auth logic
│
└─ Sub-agent 3 (DevOps)
- Loads: devops, hetzner-provisioner skills
- Context: 8K tokens
- Implements: Deployment configs
↓
All 3 work in parallel with isolated contexts
↓
Results merged back to main conversation
↓
Total effective context: 35K tokens across 3 agents
(vs 175K+ if loaded all skills in main conversation)
References
Actual Token Savings
Progressive Disclosure Savings
Scenario: User asks about Next.js
Without progressive disclosure:
Load all 35 skills: ~175,000 tokens
Context bloat: Massive
With progressive disclosure:
Metadata (all skills): ~2,625 tokens
Load relevant (2 skills): ~9,000 tokens
Total: ~11,625 tokens
Reduction: ~93%
Sub-Agent Savings
Scenario: Complex multi-domain task
Single agent approach:
Load all relevant skills: ~50,000 tokens
Main conversation history: ~80,000 tokens
Total context used: ~130,000 tokens
Risk: Approaching context limit
Sub-agent approach:
Main conversation: ~5,000 tokens (coordination only)
Sub-agent 1: ~15,000 tokens (isolated)
Sub-agent 2: ~18,000 tokens (isolated)
Sub-agent 3: ~12,000 tokens (isolated)
Total: ~50,000 tokens across 4 contexts
Reduction: ~62% (130K → 50K)
Note: Exact percentages vary by task complexity. These are approximate based on typical usage patterns.
How SpecWeave Leverages These Mechanisms
1. Skill Organization (Progressive Disclosure)
SpecWeave organizes 35+ skills with clear, focused descriptions:
# Good: Focused description
---
name: nextjs
description: NextJS 14+ App Router specialist. Server Components, SSR, routing.
---
# Bad: Vague description
---
name: frontend
description: Does frontend stuff
---
Why this matters:
- Clear descriptions help Claude identify relevance quickly
- Prevents loading irrelevant skills
- Maximizes progressive disclosure benefits
2. Agent Coordination (Sub-Agent Parallelization)
SpecWeave's role-orchestrator skill automatically:
- Detects multi-domain tasks
- Launches specialized sub-agents (PM, Architect, DevOps, etc.)
- Each sub-agent loads only its relevant skills
- Coordinates results back to main conversation
Example:
User: "/sw:inc 'Full-stack SaaS with Stripe payments'"
↓
role-orchestrator activates
↓
Launches sub-agents in parallel:
├─ PM agent (requirements)
├─ Architect agent (system design)
├─ Security agent (threat model)
└─ DevOps agent (deployment)
↓
Each loads only relevant skills in isolated context
↓
Results merged into increment spec
Common Misconceptions
❌ Myth 1: "SpecWeave has custom context manifests"
Reality: No. SpecWeave uses Claude's native progressive disclosure. Skills load based on Claude's relevance detection, not custom YAML manifests.
❌ Myth 2: "SpecWeave caches loaded context"
Reality: No custom caching. Claude Code handles caching internally (if applicable). SpecWeave doesn't implement additional caching layers.
❌ Myth 3: "70-90% token reduction"
Reality: Token savings vary by task:
- Simple tasks: 90%+ (load 1-2 skills vs all 35)
- Complex tasks: 50-70% (load 5-10 skills + use sub-agents)
- Exact percentages depend on task complexity
✅ Truth: "It just works"
Reality: Progressive disclosure and sub-agents are automatic. You don't configure them. Claude handles skill loading, sub-agent context isolation happens automatically when agents are launched.
Best Practices
For Skill Descriptions
Do:
- Be specific about what the skill does
- Include trigger keywords users might say
- List technologies/frameworks explicitly
Don't:
- Write vague descriptions ("helps with coding")
- Omit key activation triggers
- Mix multiple unrelated domains in one skill
For Sub-Agent Usage
When to use sub-agents:
- Multi-domain tasks (frontend + backend + devops)
- Parallel work (multiple features simultaneously)
- Large codebase exploration (different modules)
When NOT to use sub-agents:
- Simple single-domain tasks
- Sequential work requiring shared context
- When main conversation context is already low
Debugging Context Usage
Check Active Skills
When Claude mentions using a skill:
User: "Create a Next.js page"
Claude: "🎨 Using nextjs skill..."
This means:
- Progressive disclosure worked
- Only nextjs skill loaded (not all 35)
- Context efficient
Check Sub-Agent Usage
When Claude mentions launching agents:
Claude: "🤖 Launching 3 specialized agents in parallel..."
This means:
- Sub-agent parallelization active
- Each agent has isolated context
- Efficient multi-domain processing
Summary
SpecWeave achieves context efficiency through:
-
Progressive Disclosure (Native Claude)
- Skills load only when relevant
- Metadata-first approach
- 90%+ savings on simple tasks
-
Sub-Agent Parallelization (Native Claude Code)
- Isolated context windows
- Parallel processing
- 50-70% savings on complex tasks
No custom manifests. No custom caching. Just smart use of Claude's native capabilities.
References
GitHub Repository
Related Skills
content-collections
MetaThis skill provides a production-tested setup for Content Collections, a TypeScript-first tool that transforms Markdown/MDX files into type-safe data collections with Zod validation. Use it when building blogs, documentation sites, or content-heavy Vite + React applications to ensure type safety and automatic content validation. It covers everything from Vite plugin configuration and MDX compilation to deployment optimization and schema validation.
creating-opencode-plugins
MetaThis skill provides the structure and API specifications for creating OpenCode plugins that hook into 25+ event types like commands, files, and LSP operations. It offers implementation patterns for JavaScript/TypeScript modules that intercept and extend the AI assistant's lifecycle. Use it when you need to build event-driven plugins for monitoring, custom handling, or extending OpenCode's capabilities.
sglang
MetaSGLang is a high-performance LLM serving framework that specializes in fast, structured generation for JSON, regex, and agentic workflows using its RadixAttention prefix caching. It delivers significantly faster inference, especially for tasks with repeated prefixes, making it ideal for complex, structured outputs and multi-turn conversations. Choose SGLang over alternatives like vLLM when you need constrained decoding or are building applications with extensive prefix sharing.
evaluating-llms-harness
TestingThis Claude Skill runs the lm-evaluation-harness to benchmark LLMs across 60+ standardized academic tasks like MMLU and GSM8K. It's designed for developers to compare model quality, track training progress, or report academic results. The tool supports various backends including HuggingFace and vLLM models.
