SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillssystem-audit

system-audit

General & other

View repository

Updated 10-16 08:08

About this skill

System Audit

Scan the Claude Code configuration ecosystem to find items that can be cleaned up or optimized. Cover CLAUDE.md, MEMORY.md, and all skills in one pass.

Step 0: Scan all sources

Read CLAUDE.md

Read in order (each counts toward tokens):

  1. Global ~/.claude/CLAUDE.md
  2. Project CLAUDE.md(current directory)
  3. Project private .claude/CLAUDE.md(if it exists)

Read MEMORY.md

  1. Auto Memory — project-level .claude/projects/*/memory/MEMORY.md
  2. All .md attachments under that directory

Scan Skills

  1. Global skills — ~/.claude/skills/*/SKILL.md(including references/ subdirectories)
  2. Project skills — any subdirectory under the current directory that contains SKILL.md subdirectories

For each skill, record:

  • Number of lines in SKILL.md
  • Number of files and total lines in references/
  • Estimated token count (rough estimate: lines × 4)

ARGUMENTS filtering

If ARGUMENTS specifies a focus (claude / memory / skills), execute only the corresponding section. Default all。

Step 1: CLAUDE.md review

Analyze file by file, identifying the following issues:

1A. Redundant instructions

  • Instructions duplicated across files (global vs project-level saying the same thing)
  • The same rule expressed repeatedly within the same file

1B. Bloated content

  • Examples or explanations longer than 3 lines → can they be compressed?
  • Historical context descriptions ("we tried X before but it failed") → move to memory
  • Low-frequency instructions → convert into a skill or slash command

1C. Outdated content

  • Mentions files, features, or tools that no longer exist
  • Version numbers/dates that look outdated
  • Descriptions that contradict the existing codebase

1D. Token estimation

| 檔案 | 行數 | 預估 tokens | 問題數 |
|------|------|------------|--------|
| ~/.claude/CLAUDE.md | 200 | ~3,200 | 3 |
| CLAUDE.md | 50 | ~800 | 1 |

Step 2: MEMORY.md review

2A. Outdated entries

  • Describe features that are completed/no longer exist
  • Technical findings that contradict the existing code
  • Old records superseded by later entries

2B. Duplication with CLAUDE.md

  • memory and CLAUDE.md say the same thing → keep one place, delete the other
  • Criterion: keep high-frequency lookups in CLAUDE.md, keep low-frequency ones in memory

2C. Organization issues

  • Entries that are too long → split into a separate .md file
  • Entries with unclear categorization → suggest a category
  • Approaching the 200-line limit → mark as needing slimming down

Step 3: Skills review

3A. Overview table

| Skill | SKILL.md 行 | Refs 行 | 預估 tokens | 觸發詞 |
|-------|-------------|---------|------------|--------|
| good-writing-zh | 165 | 158 | ~1,300 | /good-writing |
| humanizer-tw | 122 | 450 | ~2,300 | humanize |
| ... | | | | |

3B. Scope overlap

Scan all skills' description + trigger words, identifying:

  • Pairs of skills with overlapping functionality (e.g., is the boundary between good-writing-zh ↔ humanizer-tw clear?)
  • Trigger word conflicts (different skills responding to the same keyword)

3C. Token efficiency

  • SKILL.md over 200 lines → suggest slimming down (extract to references/)
  • references/ over 500 lines → is there low-frequency content that can be deleted?
  • Are there skills that are actually rarely used, but expensive to load every time?

3D. Structural issues

  • Missing frontmatter (name / description / user_invocable)
  • Missing argument_hint
  • Orphan files in references/ that are not referenced by SKILL.md

Step 4: Present results + interactive execution

Produce a review report

## System Audit 報告

### Token 預算概覽
| 類別 | 每次對話載入 | 按需載入 |
|------|------------|---------|
| CLAUDE.md（全域+專案） | ~4,000 | — |
| MEMORY.md | ~3,200 | — |
| Skills（平均每次觸發） | — | ~2,000 |

### 發現問題
| # | 類別 | 嚴重度 | 描述 | 建議動作 |
|---|------|--------|------|---------|
| 1 | CLAUDE.md | 高 | 重複指令 X 出現在全域和專案級 | 刪除專案級 |
| 2 | MEMORY.md | 中 | 條目 Y 已過時 | 刪除 |
| 3 | Skills | 低 | skill Z 的 refs 有孤立檔案 | 清理 |

AskUserQuestion interaction

Use multiSelect to let the user choose which cleanup actions to execute:

  • List each finding as an option
  • Annotate severity and estimated impact (how many tokens / lines saved)
  • Only act after the user confirms

Execute cleanup

Execute the items the user selected one by one (Edit / Write). Briefly report after each action is completed.

Review principles

  • The less content loaded in each conversation, the better — CLAUDE.md and MEMORY.md are the most expensive because they are loaded every time
  • Move low-frequency content to skills or memory — instructions that are not used every time should not occupy CLAUDE.md
  • Boundaries between skills should be clear — overlapping skills waste tokens and also confuse Claude
  • Delete outdated information — incorrect memory is more dangerous than no memory

Now Execute

The user's review request is shown below ARGUMENTS:. Start from Step 0.