context-audit
Compare original and translation side by side
🇺🇸
Original
English🇨🇳
Translation
ChineseUsage Audit
使用审计
Bloated context costs more and produces worse output. This skill finds
the waste and tells you what to cut.
冗余的上下文会增加成本并降低输出质量。本skill可找出冗余部分并告诉你需要精简的内容。
Step 1: Get /context Data
步骤1：获取/context数据
Check the conversation history for /context output. If the user already
ran /context in this session, use that data. If not, ask:
"Run /context in this session terminal and let me know when you're done. I can't run
slash commands myself, but once I can see the breakdown I'll audit
everything it flags."
STOP HERE. Do NOT proceed to Step 2 until the user has ran /context. The context breakdown determines
what to audit and in what order. Without it, the audit is guessing.
Output the message above and wait for the user's next message.
查看对话历史中的/context输出。如果用户已在本次会话中执行过/context，则使用该数据。若未执行，请询问：
"请在本次会话终端中执行/context命令，完成后告知我。我无法自行执行斜杠命令，但一旦能看到上下文细分情况，我就会审计所有标记的问题。"
在此处停止。在用户执行/context之前，请勿进入步骤2。上下文细分结果决定了审计的内容和顺序。没有该数据，审计就只能靠猜测。输出上述消息并等待用户的下一条回复。
Step 2: Audit What's Bloated
步骤2：审计冗余内容
Based on the /context output, audit each category from largest to
smallest. Run checks in parallel where possible.
根据/context的输出，按从大到小的顺序审计每个类别。尽可能并行执行检查。
MCP Servers
MCP服务器
Each server loads full tool definitions into context every turn
(~15,000-20,000 tokens each).
- Count configured servers from settings.json
- Flag any with CLI alternatives (Playwright, Google Workspace, GitHub all have CLIs that cost zero tokens when idle)
- Report total MCP overhead from /context output
每个服务器每次交互都会将完整的工具定义加载到上下文中（每个约15,000-20,000 tokens）。
- 从settings.json统计已配置的服务器数量
- 标记任何有CLI替代方案的服务器（Playwright、Google Workspace、GitHub都有闲置时token消耗为零的CLI工具）
- 从/context输出中报告MCP的总开销
CLAUDE.md
CLAUDE.md
Read all CLAUDE.md files (project root, .claude/, ~/.claude/).
Count lines. Then read every rule and test against five filters:
| Filter | Flag when... |
|---|---|
| Default | Claude already does this without being told ("write clean code", "handle errors") |
| Contradiction | Conflicts with another rule in same or different file |
| Redundancy | Repeats something already covered elsewhere |
| Bandaid | Added to fix one bad output, not improve outputs generally |
| Vague | Interpreted differently every time ("be natural", "use good tone") |
If total CLAUDE.md lines > 200, check for progressive disclosure
opportunities: rules that only apply to specific tasks (API conventions,
deployment steps, testing guidelines) should move to reference files
with one-line pointers. Only recommend splitting when the file is
actually bloated -- a lean CLAUDE.md with universal context is fine
as a single file.
读取所有CLAUDE.md文件（项目根目录、.claude/、~/.claude/）。统计行数。然后逐条读取规则并通过以下五个筛选条件进行测试：
| 筛选条件 | 标记情况 |
|---|---|
| 默认能力 | Claude无需额外告知就已具备该能力（如“编写整洁代码”、“处理错误”） |
| 矛盾冲突 | 与同一文件或其他文件中的另一条规则冲突 |
| 重复冗余 | 重复了已在其他地方涵盖的内容 |
| 临时补丁 | 仅为修复某次糟糕输出而添加，并非普遍提升输出质量 |
| 表述模糊 | 每次解读都不一致（如“自然表述”、“使用良好语气”） |
如果CLAUDE.md总行数>200，检查是否有渐进式披露的机会：仅适用于特定任务的规则（API规范、部署步骤、测试指南）应移至参考文件，仅保留一行指向该文件的提示。仅当文件确实冗余时才建议拆分——内容精简、包含通用上下文的CLAUDE.md作为单个文件是没问题的。
Skills
Skills
Scan .claude/skills/*/SKILL.md. For each skill:
- Count lines (flag > 200, critical > 500)
- Run the same five filters on instructions
- Check for restated goals, hedging ("you may want to"), synonymous instructions ("be concise" + "keep it short" + "don't be verbose")
扫描.claude/skills/*/SKILL.md文件。针对每个skill：
- 统计行数（>200行标记，>500行严重标记）
- 对说明内容应用相同的五个筛选条件
- 检查是否存在重复目标、模糊表述（如“你可能想要”）、同义指令（如“简洁表述”+“保持简短”+“不要冗长”）
Settings
设置
Check settings.json for:
| Setting | Flag if | Recommended |
|---|---|---|
| autocompact_percentage_override | Missing or > 80 | 75 |
| BASH_MAX_OUTPUT_LENGTH (env) | At default (30-50K) | 150000 |
检查settings.json中的以下设置：
| 设置项 | 标记条件 | 推荐值 |
|---|---|---|
| autocompact_percentage_override | 缺失或>80 | 75 |
| BASH_MAX_OUTPUT_LENGTH（环境变量） | 为默认值（30-50K） | 150000 |
Hooks
钩子（Hooks）
Check settings.json for configuration. For each hook:
hooks- Flag hooks that run slow commands (network calls, heavy builds)
- Flag hooks with no timeout configured
- Check if any hook runs on high-frequency events (e.g., every tool call) where latency compounds
检查settings.json中的配置。针对每个钩子：
hooks- 标记执行慢命令的钩子（网络调用、重型构建）
- 标记未配置超时的钩子
- 检查是否有钩子在高频事件（如每次工具调用）中运行，此类情况会导致延迟累积
Memory Files
记忆文件
Glob . For each memory directory:
~/.claude/projects/*/memory/*.md- Count total files and estimate total lines
- Flag stale memories (check if referenced project paths still exist)
- Flag memories > 50 lines (memories should be concise pointers, not docs)
- Flag memory directories with > 20 files (index bloat)
遍历文件。针对每个记忆目录：
~/.claude/projects/*/memory/*.md- 统计总文件数并估算总行数
- 标记过期记忆（检查引用的项目路径是否仍存在）
- 标记>50行的记忆（记忆应是简洁的指向性内容，而非文档）
- 标记文件数>20的记忆目录（索引冗余）
.claudeignore
.claudeignore
Check if exists in the project root. If missing and the
project has large generated/indexed directories, recommend creating one.
Key patterns to check for:
.claudeignore- Large data directories (fixtures, seeds, dumps)
- Generated docs (typedoc, javadoc, storybook-static)
- Monorepo packages the user isn't actively working on
检查项目根目录是否存在文件。若缺失且项目包含大型生成/索引目录，建议创建该文件。需检查的关键模式：
.claudeignore- 大型数据目录（fixtures、seeds、dumps）
- 生成的文档（typedoc、javadoc、storybook-static）
- 用户未在积极开发的单仓库包
File Permissions
文件权限
Check settings.json for rules. If missing, check
whether bloat directories exist in the project:
permissions.deny| If this exists... | Should deny... |
|---|---|
| .git | .git |
| package.json | node_modules, dist, build, .next, .nuxt, .cache, coverage |
| Cargo.toml | target |
| go.mod | vendor |
| pyproject.toml / requirements.txt | pycache, .venv, *.egg-info |
检查settings.json中的规则。若缺失，检查项目中是否存在冗余目录：
permissions.deny| 若存在... | 应拒绝访问... |
|---|---|
| .git | .git |
| package.json | node_modules, dist, build, .next, .nuxt, .cache, coverage |
| Cargo.toml | target |
| go.mod | vendor |
| pyproject.toml / requirements.txt | pycache, .venv, *.egg-info |
Step 3: Score and Report
步骤3：评分与报告
Score starts at 100. Deduct per issue:
| Issue | Points |
|---|---|
| CLAUDE.md > 200 lines | -10 |
| CLAUDE.md > 500 lines | -20 |
| Per 5 rules flagged by filters | -5 |
| Contradictions between files | -10 |
| Missing autocompact override | -10 |
| Missing bash output override | -5 |
| Skill > 200 lines | -5 each |
| Skill > 500 lines | -10 each |
| Per MCP server | -3 each |
| No deny rules + bloat dirs exist | -10 |
| Hook with no timeout on frequent event | -5 each |
| Stale or oversized memory files | -5 |
| Memory directory > 20 files | -5 |
| Missing .claudeignore + large dirs exist | -5 |
Floor at 0. Output this format:
undefined初始评分为100分，根据问题扣分：
| 问题 | 扣分 |
|---|---|
| CLAUDE.md >200行 | -10 |
| CLAUDE.md >500行 | -20 |
| 每5条被筛选标记的规则 | -5 |
| 文件间存在矛盾冲突 | -10 |
| 缺失自动压缩覆盖设置 | -10 |
| 缺失bash输出覆盖设置 | -5 |
| Skill >200行 | 每个扣-5 |
| Skill >500行 | 每个扣-10 |
| 每个MCP服务器 | 每个扣-3 |
| 无拒绝规则且存在冗余目录 | -10 |
| 高频事件中运行的无超时钩子 | 每个扣-5 |
| 过期或过大的记忆文件 | -5 |
| 记忆目录文件数>20 | -5 |
| 缺失.claudeignore且存在大型目录 | -5 |
最低分为0分。输出格式如下：
undefinedUsage Audit
使用审计
Score: {N}/100 [{CLEAN|NEEDS WORK|BLOATED|CRITICAL}]
评分: {N}/100 [{CLEAN|NEEDS WORK|BLOATED|CRITICAL}]
Context Breakdown (from /context)
上下文细分（来自/context）
{Paste the key numbers from /context output}
{粘贴/context输出中的关键数据}
Issues Found
发现的问题
[{CRITICAL|WARNING|INFO}] {Category}
[{CRITICAL|WARNING|INFO}] {类别}
{What's wrong}
Fix: {One-line actionable fix}
{问题描述}
修复方案: {一行可执行的修复建议}
Rules to Cut
需移除的规则
{Each flagged rule: the text, which filter, one-line reason}
{每条被标记的规则：规则文本、所属筛选条件、一行原因说明}
Conflicts
冲突情况
{Contradictions between files, with paths}
{文件间的矛盾冲突，包含文件路径}
Top 3 Fixes
三大优先修复项
- {Highest-impact fix}
- {Second}
- {Third}
Score labels: 90-100 CLEAN, 70-89 NEEDS WORK, 50-69 BLOATED, 0-49 CRITICAL.
Severity: CRITICAL > 10pts, WARNING 5-10pts, INFO < 5pts.- {影响最大的修复方案}
- {次优修复方案}
- {第三优先修复方案}
评分标签：90-100分 → CLEAN，70-89分 → NEEDS WORK，50-69分 → BLOATED，0-49分 → CRITICAL。
严重程度：CRITICAL >10分，WARNING 5-10分，INFO <5分。Step 4: Offer to Fix
步骤4：提供修复服务
After the report:
"Want me to fix any of these? I can:
- Show you a cleaned-up CLAUDE.md with the flagged rules removed
- Add the missing settings.json configs
- Add permissions.deny rules for build artifacts
- Add timeouts to hooks missing them
- Clean up stale memory files
- Generate a .claudeignore for this project
- Show which skills to compress"
Auto-apply settings.json and permissions.deny (safe, reversible).
Show diffs for CLAUDE.md and skills -- let the user confirm before
modifying instruction files.
报告生成后：
“需要我帮你修复这些问题吗？我可以：
- 为你展示移除标记规则后的精简版CLAUDE.md
- 添加缺失的settings.json配置
- 为构建产物添加permissions.deny规则
- 为缺失超时设置的钩子添加超时
- 清理过期的记忆文件
- 为当前项目生成.claudeignore文件
- 告知你需要压缩的skills”
可自动应用settings.json和permissions.deny的修复（安全且可撤销）。对于CLAUDE.md和skills的修改，需先展示差异，获得用户确认后再修改指令文件。