context-hierarchy-design（上下文层次结构设计）

Context Hierarchy Design Skill

Design a memory hierarchy that loads context progressively based on task needs.

Purpose

Not all context is needed all the time. A well-designed hierarchy ensures agents get exactly the context they need without bloat.

When to Use

• Setting up a new project's context infrastructure
• Refactoring bloated CLAUDE.md files
• Creating task-specific context loading
• Optimizing agent startup time
• Scaling context for multiple task types

The Three-Tier Memory Strategy

Tier 1: CLAUDE.md (minimal, always loaded)
        |
        v
Tier 2: Priming Commands (task-specific, on-demand)
        |
        v
Tier 3: File Reads (just-in-time, as needed)

Design Process

Step 1: Audit Current State

Analyze existing context infrastructure:

1. Measure CLAUDE.md size
2. Count imports and their sizes
3. Review command coverage
4. Identify task types

Checklist:
- [ ] CLAUDE.md token count
- [ ] Import count and total
- [ ] Task types identified
- [ ] Command coverage mapped

Step 2: Categorize Context

For each piece of existing context, ask:

Step 3: Design Tier 1 (CLAUDE.md)

Minimal essentials only:

# Project Name

## Context
One-sentence project description.

## Tooling
- Language: X
- Runtime: Y
- Package Manager: Z

## Key Commands
- `cmd test` - Run tests
- `cmd build` - Build project

## Critical Rules
1. Most important rule
2. Second most important
3. Third most important (max 5)

Target size: less than 2KB.

Step 4: Design Tier 2 (Priming Commands)

Create commands for each task type:

Priming Command Template:

# Prime: [Task Type]

## Run
[Discovery commands for current state]

## Read
[Essential files for this task type]

## Report
[Summary prompt for understanding]

Step 5: Design Tier 3 (On-Demand)

Identify files loaded during work:

Principle: Never pre-load what can be loaded on-demand.

Output Format

Design document structure:

# Context Hierarchy Design: [Project]

## Tier 1: CLAUDE.md (Always Loaded)

**Target Size:** <2KB
**Content:**
- Project context (1 sentence)
- Tooling (language, runtime, PM)
- Key commands (3-5)
- Critical rules (3-5)

## Tier 2: Priming Commands (Task-Specific)

### /prime
**Purpose:** General codebase understanding
**Loads:** README, git status, project structure

### /prime-bug
**Purpose:** Bug fixing context
**Loads:** Recent commits, test patterns, error logs

### /prime-feature
**Purpose:** Feature development context
**Loads:** Architecture, API patterns, similar features

## Tier 3: On-Demand (Just-in-Time)

**Strategy:** Load during execution, not pre-emptively
- Source files: Read when editing
- Tests: Read when testing
- Configs: Read when configuring

## Migration Plan

1. [ ] Backup current CLAUDE.md
2. [ ] Create minimal CLAUDE.md
3. [ ] Create priming commands
4. [ ] Test with common task types
5. [ ] Iterate based on usage

Design Patterns

Pattern: Import Hierarchy

CLAUDE.md (always)
  -> imports/core.md (always, if large)
  -> imports/tooling.md (conditional)
  -> imports/workflows.md (conditional)

Pattern: Directory-Scoped Memory

/project
  CLAUDE.md (project-wide)
  /frontend
    CLAUDE.md (frontend-specific)
  /backend
    CLAUDE.md (backend-specific)

Pattern: Task-Type Priming

/prime -> Base context
/prime-{type} -> Type-specific context

Validation Criteria

After design, verify:

• CLAUDE.md under 2KB
• Every task type has priming command
• No "just in case" loading
• Progressive disclosure works
• Common workflows smooth

Anti-Patterns to Avoid

Key Quote

"Use context priming over CLAUDE.md or similar auto-loading memory files. Engineering work constantly changes, but memory files only grow."

Cross-References

• @context-priming-patterns.md - Priming command patterns
• @rd-framework.md - Reduce unnecessary context
• @context-layers.md - What loads into context
• @minimum-context-principle.md - Include only what's necessary

Version History

• v1.0.0 (2025-12-26): Initial release

Last Updated

Date: 2025-12-26 Model: claude-opus-4-5-20251101