context-management（上下文管理）

Context Management Setup

Automates context management best practices for Claude Code projects. Analyzes the current project and generates optimized configuration files.

When to Use

• Starting a new project (no CLAUDE.md exists)
• Existing project needs context optimization
• User reports context loss after compaction
• User wants to set up rules, memory, or subagent guidelines
• Before a long coding session

Workflow

Step 1: Analyze the Project

Before generating anything, read the project to understand its shape:

1. Read package.json, Cargo.toml, pyproject.toml, or equivalent to detect stack
2. Read existing CLAUDE.md if present (will update, not overwrite)
3. Check for .claude/rules/ directory
4. Scan src/ structure to identify major areas (api, components, lib, tests, etc.)
5. Check for existing tsconfig.json, .eslintrc, prettier.config for conventions
6. Read README.md if present for project context

Collect this into a mental model of the project before proceeding.

Step 2: Generate or Update CLAUDE.md

The CLAUDE.md MUST stay under 200 lines. Use @references for details.

Structure to follow:

# [Project Name]

[1-2 sentence description from README/package.json]

## Stack
[Detected stack — language, framework, DB, key libs]

## Key Commands
[Detected from package.json scripts, Makefile, etc.]
- Dev: `[command]`
- Build: `[command]`
- Test: `[command]`
- Lint: `[command]`

## Architecture
[Brief description of src/ structure — max 5-8 lines]
- @[key-file-1] for [purpose]
- @[key-file-2] for [purpose]

## Conventions
[Detected from config files — indent, quotes, semicolons, naming]

## Context Management

### Compact Instructions
When compacting, always preserve:
- List of modified files and their purpose
- Failing tests with error messages
- Architectural decisions made this session
- Current task status and next steps

When compacting, safe to discard:
- File contents that were only read for exploration
- Search results that didn't lead anywhere
- Verbose command output already acted upon

### Subagent Guidelines
- Use Explore agent for codebase searches requiring > 3 queries
- Use background agents for running test suites
- Use Plan agent before refactors touching > 5 files
- Delegate file-heavy research to subagents to protect main context

### Rules
See .claude/rules/ for scoped coding rules.

Important: If CLAUDE.md already exists, MERGE — don't replace. Add missing sections only.

Step 3: Generate Scoped Rules

Create .claude/rules/ with rules scoped to relevant paths. Only create rules for areas that actually exist in the project.

Read references/rule-templates.md for templates per area.

Rules to consider creating (only if the area exists):

Each rule file should be 10-30 lines max. Use the templates from references.

Step 4: Initialize Memory Structure

If auto-memory is enabled, create initial memory notes:

~/.claude/projects/<project-hash>/memory/MEMORY.md

Content:

# [Project Name] Memory

## Project
- Stack: [detected]
- Key paths: [detected]

## Patterns
[Leave empty — Claude fills as it learns]

## Debugging
[Leave empty — Claude fills as issues arise]

Step 5: Summary Report

After setup, show the user a summary:

Context Management Setup Complete
==================================
CLAUDE.md:     [created/updated] (X lines)
Rules:         [N] files in .claude/rules/
Memory:        [initialized/already exists]

Files created:
  - .claude/rules/code-style.md
  - .claude/rules/[others...]

Recommendations:
  - [Any project-specific suggestions]

Important Constraints

• CLAUDE.md must stay under 200 lines
• Each rule file must stay under 30 lines
• Never overwrite existing files without showing diff first
• Detect conventions from config files, don't assume
• If stack is unknown, ask the user before generating
• Use @references in CLAUDE.md instead of inlining large content