diff --git a/reader-keyword-maintenance/SKILL.md b/reader-keyword-maintenance/SKILL.md
new file mode 100644
index 0000000..aaa4cf5
--- /dev/null
+++ b/reader-keyword-maintenance/SKILL.md
@@ -0,0 +1,128 @@
+---
+name: reader-keyword-maintenance
+description: 维护 reader 项目的关键词 review 目录与低频清理 SOP。当用户要求清理或整理 reader 的关键词 review 目录、控制哪些产物长期保留、检查 suggestions JSON 是否可 apply、或执行“只保留当前 bundle / 仅保留最近少量 suggestions JSON / 删除旧 markdown”等低频维护动作时使用。不要用于正式关键词 review 生成，不要用于 reader 日报主链路。
+---
+
+# Reader Keyword Maintenance
+
+这个 skill 负责 reader 关键词治理里的**低频维护动作**，目标是把清理策略、展示稿生成、dry-run 校验和 review 目录维护从 reader 主链路里拆出来，由 OpenClaw 单独承接。
+
+## 角色边界
+
+这个 skill 负责：
+
+- 检查 `outputs/term_index/review/` 当前有哪些产物
+- 判断哪些 review 产物应长期保留、短期保留或可清理
+- 用 `apply_term_suggestions.py --dry-run` 做 apply 前校验
+- 执行低频清理和收尾动作
+
+这个 skill 不负责：
+
+- 正式生成关键词 review 建议
+- 充当 alias / stopword 正式 review 决策器
+- 替代 reader 侧的 `keyword-cleanup-review`
+
+如果用户要的是“做一轮正式关键词 review / 生成正式 suggestions JSON”，应由 OpenClaw 去调用 reader 侧 `keyword-cleanup-review`，而不是由本 skill 直接承担。
+
+## 适用场景
+
+当用户要求以下事情时使用：
+
+- 查看或解释 `reader/outputs/term_index/review/` 下有哪些文件
+- 清理旧的 review bundle、old suggestions、old markdown
+- 只保留当前 bundle 或最近少量 suggestions JSON
+- 检查 suggestions JSON 是否可被 `apply_term_suggestions.py` 消费
+- 执行一次人工 review 准备动作，但不改变 reader 主链路设计
+
+不要用于：
+
+- 日报正式生产运行
+- 生成 daily term index / term_stats 主链路
+- 自动写配置作为默认行为
+- 修改 reader 仓库里的主 SOP 作为低频维护动作的默认入口
+- 正式生成 alias / stopword / interest / watch 建议
+
+## 默认口径
+
+长期保留：
+
+- `data/term_index/daily/*.json`
+- `data/term_index/term_stats.json`
+- `configs/filter_context.personal.json`
+- `configs/term_watchlist.json`
+- `configs/term_aliases.json`
+- `configs/term_stopwords.json`
+- `configs/term_change_log.json`
+
+正式建议产物（短期保留）：
+
+- `outputs/term_index/review/term-cleanup-suggestions-YYYY-MM-DD.json`
+
+临时工作文件 / 展示层：
+
+- `outputs/term_index/review/keyword-cleanup-bundle.json`
+- `outputs/term_index/review/term-cleanup-suggestions-YYYY-MM-DD.md`
+
+更细的保留/清理判断，读取：
+
+- `references/cleanup-policy.md`
+
+## 固定 SOP
+
+### 1. 检查 review 目录
+
+- 查看 `outputs/term_index/review/` 当前有哪些文件
+- 分类成：
+  - 长期保留
+  - 短期保留
+  - 临时工作文件
+
+### 2. 准备人工审阅
+
+- 确认最新 bundle 存在
+- 确认最新 suggestions JSON 存在
+- 如需正式生成 suggestions 或 Markdown 展示稿，应转到 reader 侧 `keyword-cleanup-review`
+
+### 3. dry-run 校验 apply
+
+- 仅通过 `scripts/apply_term_suggestions.py --dry-run` 校验 suggestions JSON 是否可消费
+- 不直接落配置
+
+### 4. 清理 review 产物
+
+- 删除旧 markdown 展示稿
+- 默认只保留当前/latest bundle
+- 保守清理旧 suggestions JSON
+- 不碰 facts/state/config files
+
+## 常用动作
+
+### 1. 验证 suggestions JSON 能否被 apply 消费
+
+```bash
+cd /home/ubuntu/zhu/github/reader && \
+/usr/bin/python3.11 scripts/apply_term_suggestions.py \
+  --suggestions outputs/term_index/review/term-cleanup-suggestions-YYYY-MM-DD.json \
+  --accept-interest "Claude Code" \
+  --dry-run
+```
+
+## 清理策略
+
+保守策略：
+
+- bundle：默认只保留当前最新一份
+- markdown：默认不归档，需要时现生成
+- suggestions JSON：只保留最近少量几份或已应用过的记录
+
+## 维护原则
+
+- 优先减少中间产物，不要让 review 工作文件变成长期资产
+- JSON suggestions 是 review / apply 之间唯一正式输入
+- Markdown 只是展示层，不是系统真相
+- 任何清理动作前，先确认不会影响当前待审阅或待 apply 的 suggestions JSON
+- 正式关键词 review 的生成入口不在这里，而在 reader 侧 `keyword-cleanup-review`
+
+执行清理或 review-prep 前，先读取：
+
+- `references/maintenance-checklist.md`
diff --git a/reader-keyword-maintenance/references/alias-review-example.md b/reader-keyword-maintenance/references/alias-review-example.md
new file mode 100644
index 0000000..acb590d
--- /dev/null
+++ b/reader-keyword-maintenance/references/alias-review-example.md
@@ -0,0 +1,33 @@
+# Alias Review Example
+
+Use this as the default style reference when preparing a low-frequency alias review report.
+
+## Suggested merges
+
+- `Claude code -> Claude Code`
+  - 理由：明显属于同一产品名，仅是大小写写法不一致。
+  - 证据：`Claude Code` 在最近多日持续出现，而小写写法只是在少量上下文中作为变体出现。
+
+- `Sub-Agent -> SubAgent`
+  - 理由：更像词形差异，不构成新的独立概念。
+  - 证据：两者都围绕同一 agent 架构语境出现，且没有稳定区分语义。
+
+## Not recommended to merge
+
+- `Skills ↔ Agent Skills`
+  - 理由：前者过泛，后者更具体，当前强行归并会损失粒度。
+
+- `Anthropic ↔ Claude`
+  - 理由：公司名与产品名并不等价，不应直接视为一个关键词。
+
+## Needs human judgment
+
+- `AI助手 ↔ AI Agent`
+  - 风险点：语义可能接近，但中文表述范围更宽，是否并入需要结合实际使用语境判断。
+
+## Style notes
+
+- Keep the report concise.
+- Give judgment first, then evidence.
+- Do not claim anything is already applied.
+- Prefer conservative proposals over broad semantic merging.
diff --git a/reader-keyword-maintenance/references/cleanup-policy.md b/reader-keyword-maintenance/references/cleanup-policy.md
new file mode 100644
index 0000000..0cbb527
--- /dev/null
+++ b/reader-keyword-maintenance/references/cleanup-policy.md
@@ -0,0 +1,106 @@
+# Cleanup Policy
+
+## Purpose
+
+This reference defines how `reader-keyword-maintenance` should treat keyword governance artifacts.
+
+The goal is to keep long-term assets small and stable while allowing review-time working files to exist when needed.
+
+This policy is for low-frequency maintenance only.
+It does not replace the formal keyword review generation flow in reader.
+
+## Artifact classes
+
+### Long-term assets
+
+Keep these by default:
+
+- `data/term_index/daily/*.json`
+- `data/term_index/term_stats.json`
+- `configs/filter_context.personal.json`
+- `configs/term_watchlist.json`
+- `configs/term_aliases.json`
+- `configs/term_stopwords.json`
+- `configs/term_change_log.json`
+
+These are facts or active state.
+
+### Short-term decision artifacts
+
+Keep selectively:
+
+- `outputs/term_index/review/term-cleanup-suggestions-YYYY-MM-DD.json`
+
+Recommended policy:
+
+- keep only recent few files, or
+- keep only files that were actually used for apply decisions
+
+### Temporary working/display files
+
+Treat as disposable unless the user explicitly asks to archive them:
+
+- `outputs/term_index/review/keyword-cleanup-bundle.json`
+- `outputs/term_index/review/term-cleanup-suggestions-YYYY-MM-DD.md`
+
+Recommended policy:
+
+- bundle: keep only latest current file
+- markdown: generate on demand, do not archive by default
+
+## Default maintenance actions
+
+### Safe checks
+
+Before deleting anything:
+
+1. confirm the target is not the current bundle under active review
+2. confirm the target suggestions JSON is not the one about to be applied
+3. never delete configs or facts during review cleanup
+
+### Safe cleanup order
+
+1. remove or overwrite old markdown display drafts
+2. keep only latest bundle file
+3. prune old suggestions JSON files conservatively
+
+## Decision rules
+
+### When user says "clean review artifacts"
+
+Default action:
+
+- keep facts/state untouched
+- keep current suggestions JSON
+- remove markdown drafts if they are old and clearly derived display files
+- keep bundle only as current working file
+
+### When user says "prepare manual review"
+
+Default action:
+
+- ensure latest bundle exists
+- ensure latest suggestions JSON exists
+- do not generate formal suggestions or markdown here by default
+- if the user wants formal review material, route to reader-side `keyword-cleanup-review`
+
+### When user says "review alias candidates" or "review stopword candidates"
+
+Default action:
+
+- explain that formal keyword review generation belongs to reader-side `keyword-cleanup-review`
+- keep this skill focused on maintenance, cleanup, display generation, and dry-run validation
+- only proceed here if the user explicitly wants a low-frequency maintenance view rather than the formal review flow
+
+### When user says "what can be deleted"
+
+Explain in three buckets:
+
+- must keep
+- can keep temporarily
+- safe to regenerate/delete
+
+## Non-goals
+
+This policy does not change reader production logic.
+It only governs low-frequency maintenance and cleanup decisions in OpenClaw.
diff --git a/reader-keyword-maintenance/references/maintenance-checklist.md b/reader-keyword-maintenance/references/maintenance-checklist.md
new file mode 100644
index 0000000..1dc12d6
--- /dev/null
+++ b/reader-keyword-maintenance/references/maintenance-checklist.md
@@ -0,0 +1,55 @@
+# Maintenance Checklist
+
+Use this checklist before doing any cleanup or review-maintenance action for reader keyword artifacts.
+
+## Pre-check
+
+1. Confirm the current repo root is `/home/ubuntu/zhu/github/reader`
+2. Confirm the user asked for a maintenance / cleanup / review-prep action
+3. If the user actually wants formal keyword review generation, route to reader-side `keyword-cleanup-review` instead of using this maintenance skill
+4. Identify whether the action targets:
+   - current bundle
+   - suggestions JSON
+   - markdown display draft
+   - old review outputs
+
+## Safety check
+
+Before deleting or overwriting anything:
+
+1. Do not touch:
+   - `data/term_index/daily/*.json`
+   - `data/term_index/term_stats.json`
+   - `configs/filter_context.personal.json`
+   - `configs/term_watchlist.json`
+   - `configs/term_aliases.json`
+   - `configs/term_stopwords.json`
+   - `configs/term_change_log.json`
+2. Confirm the suggestions JSON to keep is not the one about to be applied
+3. Treat markdown drafts as disposable only after confirming they are display-only artifacts
+
+## Review-prep flow
+
+When preparing manual review:
+
+1. Ensure the latest bundle exists
+2. Ensure the latest suggestions JSON exists
+3. Generate markdown only if the user explicitly wants human-readable review material
+4. Prefer showing conclusions in chat before creating more files
+
+## Cleanup flow
+
+Recommended order:
+
+1. remove old markdown display drafts
+2. keep only the current/latest bundle
+3. prune old suggestions JSON conservatively
+4. leave facts/state/config untouched
+
+## Post-check
+
+After the action:
+
+1. verify the expected kept files still exist
+2. verify no config file was accidentally changed
+3. summarize what was kept vs removed