English Changelog

• Design: Major upgrade incorporating Impeccable, GetDesign, and Kora reference material; sharpens aesthetic direction and removes generic defaults.
• Design: Refined craft details and skill contracts based on real-world usage review; tighter guidance on tier detection and output format.
• Health: Added Part C context-effectiveness audit to the inspector-context agent for deeper conversation-quality analysis.
• Health: Replaced head -c byte-slicing with line-safe JSONL sampling; tightened correction-signal detection to cut false positives.
• Write: Distilled top friction patterns from 30-day usage; added concrete counter-examples for bridge-sentence and TL;DR rules.
• Statusline: Added rate-limits cache fallback so token counters survive account switches; fixed cache-write guard to prevent stale data overwrites.
• Verification: Introduced Makefile with smoke tests for statusline and health, replacing scattered inline shell checks.

中文更新日志

• 设计技能: 全面升级，融入 Impeccable、GetDesign 和 Kora 参考素材，审美方向更清晰，彻底去掉通用默认风格。
• 设计技能: 根据实际使用反馈精炼工艺细节和技能契约，层级检测和输出格式指导更严谨。
• Health: inspector-context agent 新增 Part C 上下文效果审计，对话质量分析更深入。
• Health: 用逐行安全的 JSONL 采样替换 head -c 字节截断；收紧 correction 信号检测，减少误报。
• 写作技能: 提炼 30 天使用中的高频摩擦点，为桥接句和 TL;DR 规则补充具体反例。
• Statusline: 新增 rate-limits 缓存回退，切换账号后 token 计数不再丢失；修复缓存写入条件，防止陈旧数据覆写。
• 验证: 新增 Makefile，包含 statusline 和 health 的 smoke test，取代分散的内联 shell 检查。

Update: npx skills add tw93/Waza@latest · ⭐ tw93/Waza

Changelog

1. Write: Editing pass now catches paragraph-end summary sentences first and stops over-correcting the author's natural voice.
2. Learn: When two sources contradict on a fact, both positions are now recorded rather than one being silently picked.
3. Hunt: Visual and rendering bugs now follow a static-analysis-first protocol: inspect paint layers in DevTools before adding logs.
4. Check: PR comment guidance now requires natural paragraphs and thanks the contributor first, not bullet-point reports.
5. Read: Fetch cascade retries each proxy once before falling back; GitHub file URLs now route to raw.githubusercontent.com or gh api instead of the proxy.

更新日志

1. Write: 改稿优先清除段末总结句，不再过度去口语化，英文清单补 emoji 删除。
2. Learn: 来源互相矛盾时，现在记录双方立场和依据，不再默默选一个。
3. Hunt: 视觉/渲染 bug 现在先走静态分析流程，在 DevTools 检查绘制层，再加日志。
4. Check: PR 评论规范改为自然段落，先感谢贡献者，不用列表汇报格式。
5. Read: 抓取级联现在对每个代理各重试一次；GitHub 文件链接改为走 raw.githubusercontent.com 或 gh api。

https://github.com/tw93/Waza

Changelog

1. /read and /learn: now support web search plugins (e.g., PipeLLM) for better source discovery and JavaScript-rendered pages.
2. /think: added TDD protocol enforcing red-green-refactor cycles after design approval.
3. Agent discovery: added .agents/skills/ symlinks for agents that auto-discover skills from that path.

更新日志

1. /read 和 /learn: 现在支持网页搜索插件（如 PipeLLM），提升来源发现和 JavaScript 渲染页面支持。
2. /think: 新增 TDD 协议，设计通过后强制执行红-绿-重构循环。
3. Agent 发现: 新增 .agents/skills/ 符号链接，支持从该路径自动发现技能的 agent。

Update: npx skills add tw93/Waza@latest · ⭐ https://github.com/tw93/Waza

English Changelog

• Skill philosophy: replaced prescriptive step-by-step instructions in health, check, think, and hunt with goal statements, letting the model choose the best approach as capabilities improve.
• Health refactor: extracted the 180-line data-collection bash script into scripts/collect-data.sh, reducing health SKILL.md by 60%.
• Check refactor: removed hardcoded git and gh commands; trimmed GitHub Operations to constraint-only rules.
• Read: defuddle.md is now the primary fetch proxy; r.jina.ai falls back if it fails.
• Statusline: setup script now detects an existing statusline and prompts before overwriting (default Y).
• Rename: health agents renamed to inspector-context/inspector-control to match the reviewer-* pattern; verify.sh renamed to run-tests.sh; english-coaching.md renamed to coaching-en.md.

中文更新日志

• 技能哲学: 将 health、check、think、hunt 中逐步指令替换为目标陈述，让模型随能力提升自主选择最佳方式。
• Health 重构: 将 180 行数据采集 bash 脚本提取为 scripts/collect-data.sh，health SKILL.md 体积缩减 60%。
• Check 重构: 移除硬编码的 git 和 gh 命令；GitHub Operations 精简为纯约束规则。
• Read: defuddle.md 升为首选抓取代理，r.jina.ai 作为备用。
• Statusline: 安装脚本现在检测已有 statusline 并在覆盖前提示确认（默认 Y）。
• 重命名: health agents 改为 inspector-context/inspector-control，与 reviewer-* 命名风格一致；verify.sh 改为 run-tests.sh；english-coaching.md 改为 coaching-en.md。

Update: npx skills add tw93/Waza@latest  ·  ⭐ github.com/tw93/Waza

English Changelog

• Version Unification: all 8 skills moved to a single version 3.0.0 for collection-level consistency.
• Rationalization Watch: /hunt now includes a named table of 5 diagnostic-failure phrases ("I'll just try this one thing", "I'm confident it's X", etc.) with rebuttals that redirect back to evidence-based diagnosis.
• Condition-Based Waiting: /hunt Known Failure Shapes gains a new row for replacing fixed sleeps with condition polling (file exists, port responds, process exits).
• Verification Anti-Patterns: /check Verification section now lists 5 specific rationalization phrases ("should work now", "probably fixed", "I'm confident", etc.) as hard stops before claiming done.
• No-Placeholder Rule: /think Phase 3 now explicitly forbids TBD, TODO, "implement later", and similar placeholders in approved plans.

中文更新日志

• 版本统一: 全部 8 个 skill 统一升至 3.0.0，告别各自为政的版本号。
• 自我合理化警示表: /hunt 新增 5 条诊断失败的惯用借口（"就试这一个小改动"、"我有把握是 X"等），每条附上应对规则，防止在没有假设的情况下随机修改代码。
• 条件等待模式: /hunt 已知故障形态表新增"条件等待"行，替代固定时间 sleep，改为轮询状态（文件存在、端口响应、进程退出）。
• 验证反模式清单: /check 验证节点列出 5 个合理化短语（"应该好了"、"感觉没问题"等），出现任何一个即强制先跑验证命令再声明完成。
• 禁止占位符: /think Phase 3 明确禁止在已批准的计划中出现 TBD、TODO、"以后实现"等占位符。

Update: npx skills add tw93/Waza@latest · ⭐ Star on GitHub

Why Waza

This started as claude-health: a single skill to audit Claude Code configuration. It worked well, and I kept adding more skills for the habits I use every day — debugging, design, writing, research, code review.

At some point I looked at tools like Superpowers and gstack. They are impressive, but they are heavy. Too many skills, too much setup, built for power users. Not the right fit for a younger engineer who just wants a clean, focused workflow.

So instead of growing claude-health into another heavy toolkit, I gave it a new name: Waza (技), the Japanese term for a technique practiced until it becomes instinct. Eight focused skills. Each one does one thing well. Nothing unnecessary.

English Changelog

• Rename: claude-health is now Waza, a full skill collection for the complete engineer.
• Skills: eight skills covering the core engineering workflow: /think, /design, /check, /hunt, /write, /learn, /read, /health.
• English Coaching: passive grammar correction shipped as a template, not a skill. One curl command adds it to any Claude Code setup.
• Progressive disclosure: large reference content extracted to supporting files for /english, /design, /read — SKILL.md stays thin and fast to load.
• AI usability: explicit AskUserQuestion gates, concrete verification commands, fixed placeholder text, and correct allowed-tools declarations across all skills.
• Descriptions: all descriptions trimmed to trigger-focused format, reducing constant context overhead.

中文更新日志

• 重命名: claude-health 正式升级为 Waza，覆盖完整工程工作流的 skill 合集。
• Skills: 八个技能覆盖核心工程习惯：/think、/design、/check、/hunt、/write、/learn、/read、/health。
• English Coaching: 被动英文纠错以模板形式发布，不再是 skill。一行 curl 命令即可添加到任意 Claude Code 配置。
• 渐进式加载: /design、/read 等 skill 的大型参考内容拆分到 references/ 子目录，SKILL.md 保持精简。
• AI 可用性: 所有 skill 补充了明确的 AskUserQuestion 节点、具体的验证命令、修正了占位符文本和 allowed-tools 声明。
• 描述精简: 所有描述改为触发条件格式，减少常驻上下文开销。

Update: npx skills add tw93/Waza@latest -g -y · ⭐ Star on GitHub

• Security: Added credential redaction instruction before pasting Step 1 data to subagents, addressing Snyk W007 credential handling finding.
• Security: Added untrusted-input boundary declaration to Agent 1, mitigating prompt injection risk when processing external skill content.

• 安全: 在向 subagent 粘贴 Step 1 数据前加入 credential 脱敏说明，修复 Snyk W007 凭证处理问题。
• 安全: 在 Agent 1 开头加入不可信输入边界声明，缓解处理外部 skill 内容时的 prompt injection 风险。

Update: npx skills add tw93/claude-health@latest

⭐ tw93/claude-health

• Security: Removed embedded bash grep scan that caused false positives in Socket, Snyk, and Gen scanners; Agent 1 already covers the same checks via model analysis of skill content.

• 安全: 移除了内嵌的 bash grep 扫描块，该模块导致 Socket、Snyk、Gen 安全扫描器误报；Agent 1 通过对 skill 内容的模型分析已覆盖相同的安全检查。

Update: npx skills add tw93/claude-health@latest

⭐ tw93/claude-health

English Changelog

• Invocation: remove disable-model-invocation from the health skill so explicit Claude Code invocation is no longer blocked by the current plugin-skill behavior.
• Gitignore: switch the credential ignore check to git check-ignore -v and only treat repo-local .gitignore rules as a passing safeguard.
• Runtime: speed up Step 1 data collection by preferring indexed git scans, merging settings parsing, and capping large conversation and denial log reads.
• Audit: add MCP filesystem allowlist detection, safer handling for missing skill directories and the active session file, and lighter checks for config-focused projects.
• Prompts: split agent prompts, tighten health skill instructions, improve report presentation, and align output language and progress lines with communication rules.

中文更新日志

• 调用: 移除 health skill 的 disable-model-invocation，避免 Claude Code 当前 plugin skill 行为继续拦截显式调用。
• Gitignore: 将凭据忽略检查改为 git check-ignore -v，并且只有仓库内 .gitignore 规则才视为通过。
• 运行时: 通过优先使用 git 索引扫描、合并 settings 解析、限制大体积会话和 denial 日志读取，加快 Step 1 数据收集。
• 审查: 新增 MCP filesystem allowlist 检测，补强缺失 skill 目录和活跃会话文件的处理，并对配置型项目放宽不必要的审查项。
• 提示词: 拆分 agent 提示，收紧 health skill 指令，改进报告展示，并让输出语言与进度文案遵循 communication 规则。

Update: npx skills add tw93/claude-health@latest -a claude-code -s health -g -y

⭐ github.com/tw93/claude-health

English Changelog

• Runtime: cap SKILL FULL CONTENT to 5 skills × 80 lines to prevent indefinite hangs on large installs.
• Install: README install command now explicitly passes -a claude-code -s health to ensure correct placement.
• Troubleshooting: new section covering Unknown skill, wrong install path, long-running audit, and scope questions.
• Audit: add hook error-handling check, CLAUDE.md conflict detection, MEMORY.md urgency weighting, slow-test and output-truncation checks, startup context budget estimate, nested CLAUDE.md detection, and gitignore credential check.
• Fix: loosen allowedTools hygiene to only flag genuinely dangerous patterns; add fallback messages for silent-failure sections; fix project path sed to match Claude Code directory naming.
• UX: merge bash blocks into one to reduce per-command permission prompts; inline all data for agents so they need no extra file reads.
• Metadata: correct owner name and email in marketplace.json; add disable-model-invocation flag to skill frontmatter.

中文更新日志

• 运行时: SKILL FULL CONTENT 采样上限改为 5 个 skill × 80 行，修复大量 skill 时 /health 卡死问题。
• 安装: README 安装命令补全 -a claude-code -s health，确保文件落到正确路径。
• 问题排查: 新增 Troubleshooting 章节，覆盖 Unknown skill、路径错误、运行超时、scope 说明四类常见问题。
• 审查项: 新增 hook 错误处理检查、CLAUDE.md 冲突检测、MEMORY.md 紧急权重、慢测试和输出截断检查、启动上下文预算估算、嵌套 CLAUDE.md 检测、gitignore 凭据检查。
• 修复: 放宽 allowedTools 卫生检查，只标记真正危险的模式；为 silent-failure 段落补充兜底提示；修复项目路径 sed 表达式以匹配 Claude Code 目录命名规则。
• 体验: 将多个 bash 块合并为一个，减少逐条授权提示；所有数据内联传递给子 agent，无需额外文件读取。
• 元数据: 修正 marketplace.json 中 owner 信息；skill frontmatter 加入 disable-model-invocation 标志。

Update: npx skills add tw93/claude-health@latest -a claude-code -s health -g -y

⭐ github.com/tw93/claude-health