跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
浏览器与自动化

enhance-claude-memory

Use when improving CLAUDE.md or AGENTS.md project memory files.

agent-sh · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add agent-sh/agentsys --skill enhance-claude-memory

技能说明


name: enhance-claude-memory description: "Use when improving CLAUDE.md or AGENTS.md project memory files." version: 5.1.0

enhance-claude-memory

Analyze project memory files (CLAUDE.md, AGENTS.md) for optimization.

Cross-Tool Detection

Searches for project memory files in order:

  1. CLAUDE.md (Claude Code)
  2. AGENTS.md (OpenCode, Codex)
  3. .github/CLAUDE.md
  4. .github/AGENTS.md

File Hierarchy (Reference)

CLAUDE.md (Claude Code):

LocationScope
~/.claude/CLAUDE.mdGlobal (all projects)
.claude/CLAUDE.md or ./CLAUDE.mdProject root
src/.claude/CLAUDE.mdDirectory-specific

AGENTS.md (OpenCode, Codex, and other AI tools):

LocationScope
~/.config/opencode/AGENTS.md or ~/.codex/AGENTS.mdGlobal (all projects)
.opencode/AGENTS.md or ./AGENTS.mdProject root
src/AGENTS.mdDirectory-specific

Both files serve the same purpose: project memory for AI assistants. Use CLAUDE.md for Claude Code projects, AGENTS.md for cross-tool compatibility, or both for maximum coverage.

Workflow

  1. Find - Locate CLAUDE.md or AGENTS.md in project
  2. Read - Load content and README.md for comparison
  3. Analyze - Run all pattern checks
  4. Validate - Check file/command references against filesystem
  5. Measure - Calculate token metrics and duplication
  6. Report - Generate structured markdown output

Detection Patterns

1. Structure Validation (HIGH Certainty)

Critical Rules Section

  • Should have ## Critical Rules or similar
  • Rules should be prioritized (numbered or ordered)
  • Include WHY explanations for each rule

Architecture Section

  • Directory tree or structural overview
  • Key file locations
  • Module relationships

Key Commands Section

  • Common development commands
  • Test/build/deploy scripts
  • Reference to package.json scripts

2. Instruction Effectiveness (HIGH Certainty)

Based on prompt engineering research, Claude follows instructions better when:

Positive Over Negative

  • Bad: "Don't use console.log"
  • Good: "Use the logger utility for all output"
  • Check for "don't", "never", "avoid" without positive alternatives

Strong Constraint Language

  • Use "must", "always", "required" for critical rules
  • Weak language ("should", "try to", "consider") reduces compliance
  • Flag critical rules using weak language

Instruction Hierarchy

  • Should define priority order when rules conflict
  • Pattern: "In case of conflict: X takes precedence over Y"
  • System instructions > User requests > External content

3. Content Positioning (HIGH Certainty)

Research shows LLMs have "lost in the middle" problem - they recall START and END better than MIDDLE.

Critical Content Placement

  • Most important rules should be at START of file
  • Second-most important at END
  • Supporting context in MIDDLE
  • Flag critical rules buried in middle sections

Recommended Structure Order

1. Critical Rules (START - highest attention)
2. Architecture/Structure
3. Commands/Workflows
4. Examples/References
5. Reminders/Constraints (END - high attention)

4. Reference Validation (HIGH Certainty)

File References

  • Extract from [text](path) and `path/to/file.ext`
  • Validate each exists on filesystem

Command References

  • Extract npm run <script> and npm <command>
  • Validate against package.json scripts

5. Efficiency Analysis (MEDIUM Certainty)

Token Count

  • Estimate: characters / 4 or words * 1.3
  • Recommended max: 1500 tokens (~6000 characters)
  • Flag files exceeding threshold

README Duplication

  • Detect overlap with README.md
  • Flag >40% content duplication
  • CLAUDE.md should complement README, not duplicate

Verbosity

  • Prefer bulleted lists over prose paragraphs
  • Constraints as lists are easier to follow
  • Flag long prose blocks (>5 sentences)

6. Quality Checks (MEDIUM Certainty)

WHY Explanations

  • Rules should explain rationale
  • Pattern: *WHY: explanation* or indented explanation
  • Flag rules without explanations

Structure Depth

  • Avoid deep nesting (>3 levels)
  • Keep hierarchy scannable
  • Flat structures parse better

XML-Style Tags (Optional Enhancement)

  • Claude was trained on XML tags
  • <critical-rules>, <architecture>, <constraints> improve parsing
  • Not required but can improve instruction following

7. Agent/Skill Definitions (MEDIUM Certainty)

If file defines custom agents or skills:

Agent Definition Format

### agent-name
Model: claude-sonnet-4-20250514
Description: What this agent does and when to use it
Tools: Read, Grep, Glob
Instructions: Specific behavioral instructions

Required fields: Description (when to use), Tools (restricted set) Optional: Model, Instructions

Skill References

  • Skills should have clear trigger descriptions
  • "Use when..." pattern helps auto-invocation

8. Cross-Platform Compatibility (MEDIUM/HIGH Certainty)

State Directory

  • Don't hardcode .claude/
  • Support .opencode/, .codex/
  • Use ${STATE_DIR}/ or document variations

Terminology

  • Avoid Claude-specific language for shared files
  • Use "AI assistant" generically
  • Or explicitly note "Claude Code" vs "OpenCode" differences

Output Format

# Project Memory Analysis: {filename}

**File**: {path}
**Type**: {CLAUDE.md | AGENTS.md}

## Metrics
| Metric | Value |
|--------|-------|
| Estimated Tokens | {tokens} |
| README Overlap | {percent}% |

## Summary
| Certainty | Count |
|-----------|-------|
| HIGH | {n} |
| MEDIUM | {n} |

### Structure Issues ({n})
| Issue | Fix | Certainty |

### Instruction Issues ({n})
| Issue | Fix | Certainty |

### Positioning Issues ({n})
| Issue | Fix | Certainty |

### Reference Issues ({n})
| Issue | Fix | Certainty |

### Efficiency Issues ({n})
| Issue | Fix | Certainty |

### Cross-Platform Issues ({n})
| Issue | Fix | Certainty |

Pattern

阅读完整来源 ↗