Khazix Skills: A Personal Agent Skills Collection With Strong Opinions and Guardrails

What This Collection Actually Is

Khazix Skills is a set of six Agent Skills, each a structured instruction set that an AI agent can load directly. The author states these are skills they use daily in their own projects and open sourced them after they proved useful. The target audience is people who use agent tools like Claude Code, Codex, Qoder, Kimi Code, iFlow, CodeBuddy or Cursor, which support the Agent Skills open standard. The skills are not a framework or a library. They are prompt-level protocols written as SKILL.md files. The repository is in Chinese with an English README available, and the primary language is listed as Python, though the skills themselves are mostly instruction documents rather than code. The license is MIT, which means you can reuse the skill definitions freely, but the README gives no information about contribution guidelines or release cadence.

The Leader Skill: Goals With a Harness

The leader skill addresses a specific failure mode: when an AI agent runs autonomously for hours, a poorly defined goal causes wasted effort in the wrong direction. The author's argument is that the most important part of a goal is not what to do, but what not to do. They cite the Apollo example: the key was not landing on the moon, but bringing the crew back safely. The skill produces a goal task document, which you can paste into a goal mode like Claude Code's /goal or Codex's goal mode. The author describes a process of seven questions covering purpose, completion state, evidence, anti-cheating, boundaries, trade-offs and unknowns. A notable detail is the zero-th question: the skill first explores the codebase and runs commands itself, because documentation may reference commands that do not exist. The output is pure Markdown, so it works even without a goal mode. The README suggests using a strong planning model for the goal and a strong execution model for the long run, naming specific models like Claude Fable 5 and GPT-5.6 Sol, which may become outdated quickly.

Storage Analyzer: Read Only, With a Local Web UI

The storage-analyzer skill scans an entire Mac or Windows disk and produces an interactive HTML report in the browser. The report shows disk overview, top five space consumers, cleanup priority and a three color classification: green for pure cache, yellow for user data, red for system or running application data. The core rule is that scanning is strictly read only. The agent never deletes anything by itself. Deletion requires you to click a button in the browser and then confirm in a second dialog. The local service runs on 127.0.0.1 with a random port and a token, and the permission model has three whitelist tiers: green can delete, yellow can only move to trash, red can only open the folder. The README claims macOS is fully tested, while Windows code is ready but the author advises caution on first use. A concrete example given is a 97 GB UUID Container, which the skill identifies as Bilibili offline video cache and recommends clearing within the Bilibili client rather than manually deleting. This is a genuinely useful distinction compared to a rule based cleaner like CleanMyMac, which would just label it as user cache.

Neat-Freak: Fighting Documentation Rot

The neat-freak skill, triggered by /neat, aligns project documentation, CLAUDE.md or AGENTS.md files, and agent memory after a task. The author's diagnosis is that documentation becomes stale as code iterates, and agents then operate on outdated assumptions, making them progressively less useful. The skill works on three layers with different audiences: CLAUDE.md and AGENTS.md for the current AI, docs and README for humans, and the agent's own memory for cross session continuity. It also audits rules as knowledge, checking whether CLAUDE.md and AGENTS.md are consistent, whether required files exist, and whether referenced paths still exist. Version 3.0 adds a lightweight path for small projects without git or rule files: it aligns the README to the current code, creates a minimal AI rules file, and lists session leftovers like PLAN.md, debug scripts and xxx_old files for confirmation. The hard rule is that it never deletes anything on its own, only produces a candidate list for you to approve. Machine generated memory is read only by default, and reading a command in a file is not treated as authorization. The README explicitly says it will not trigger for pure code tasks or data整理 or weekly report requests, only for project knowledge wrap up.

Getting Running: Installation and Trigger Phrases

Installation is straightforward for any tool that supports Agent Skills. You say to the agent: install this skill: https://github.com/KKKKhazix/khazix-skills/tree/main/<skill-name>, replacing <skill-name> with the skill directory name like neat-freak or hv-analysis. The agent clones the repository to the appropriate directory. If your agent does not support Skills, the README says you can download the SKILL.md file and use it as a project rules file or paste it directly into the conversation. Each skill has trigger phrases in Chinese and some in English. For storage-analyzer, examples include 帮我看看存储, C 盘满了, and storage analysis. For leader, triggers include 帮我给 agent 写个目标 and 让 agent 自己跑这个项目. For aihot, you can say 今天 AI 圈有什么新东西 or 最近 OpenAI 有什么发布. The aihot skill requires no API key and no MCP server, pulling daily AI HOT reports from aihot.virxact.com. It supports date specific reports, hot events sorted by heat, category filters like model, product, industry, paper or technique, and time windows of 24 hours or 7 days, plus keyword search and local synchronization of current selections.

Writing and Research: Two Opinionated Skills

The hv-analysis skill produces a 10,000 to 30,000 character PDF research report on a product, company, concept or person. It runs two tracks in parallel: a vertical track that traces the subject from birth to present, and a horizontal track that compares all major competitors in the same period. The intersection of these two lines is meant to reveal insights not visible from either alone. The README is honest about its limitations: it is not for simple term definitions, which would be overkill, and not for writing WeChat articles, which is the job of khazix-writer. The khazix-writer skill replicates the author's personal writing style for their public account 数字生命卡兹克. It has a strong stance and will refuse to write certain clichés like 赋能, 抓手, 闭环, 首先...其次, 在当今 AI 快速发展的时代, and 说白了. This is a double edged sword: it gives you a distinctive voice, but only if you actually like the author's style. The README warns that if you want generic good writing, this skill is not for you. Both skills are triggered by simple phrases like 研究一下 Cursor 这家公司 or 帮我做个 deep research for hv-analysis, and by providing a PDF, voice transcript or news link for khazix-writer.

Limitations and the Wrong Tool Cases

The most obvious limitation is that the entire collection is built around the author's personal workflow and Chinese language context. Trigger phrases are mostly Chinese, and the writing skill explicitly rejects common Chinese bureaucratic phrases, which is fine for that audience but irrelevant or counterproductive for English writing. The README gives no information about testing coverage for the aihot skill or the leader skill beyond the author's own use. The storage-analyzer skill is the only one with a stated platform testing status, and Windows is explicitly marked as less proven. There are no releases, no version numbers for the collection, and no changelog, so you cannot track what changed between versions. The neat-freak skill's claim that it audits whether rules were actually executed is ambitious, and the README does not explain how it verifies execution, only that it checks consistency of files and paths. The leader skill's output depends heavily on the quality of the model you use, and the specific model names mentioned will age poorly. For the wrong tool case, hv-analysis explicitly says it is not for simple queries, and khazix-writer is not for generic style. The aihot skill depends on an external service, aihot.virxact.com, which could change or disappear, and the README does not specify what happens if that service is down.

Maintenance, Licensing and Alternatives

The repository has no recent releases and no homepage, and the last push date is unknown, so you cannot infer maintenance activity from the README. The MIT license means you can modify and redistribute the skills freely, but you should check each SKILL.md for any embedded third party content or external service terms. The aihot skill relies on an external API, so its maintenance cost includes tracking changes to that endpoint. The storage-analyzer runs a local web server, which adds a security surface, though the README describes a reasonable token and whitelist model. As an alternative, consider the built in goal mode in Claude Code or Codex, which provides the execution loop but not the structured goal definition process; leader is a layer on top of that. For disk cleanup, CleanMyMac is the named commercial alternative, but it is rule based and does not explain what each item is, which is the core difference. For research reports, any general deep research tool could produce similar output, but hv-analysis's value is the explicit two axis method. For writing, the alternative is to craft your own style guide, but khazix-writer gives you a ready made, opinionated one. Before adopting, read the SKILL.md files directly, because the README is a marketing summary and the actual instructions may differ in detail.

Editorial conclusion

Adopt Khazix Skills if you use an Agent Skills compatible tool like Claude Code or Codex and you want structured, opinionated workflows for defining goals, cleaning up project knowledge, or scanning disk usage. The leader and neat-freak skills are the most distinctive because they encode hard rules about not doing harm: no deletions without confirmation, no unconstrained goals. Skip it if you need generic writing style or a simple dictionary lookup, as hv-analysis explicitly warns against that. Before adopting, verify that your agent supports SKILL.md loading, and test the storage-analyzer on a non-critical machine first, especially on Windows where the README says code is ready but only macOS is fully tested. The repository has no releases and no homepage, so check the commit history and individual SKILL.md files for maintenance signals before relying on it in production.

Official sources

1. Official README
2. Project repository