khazix-skills: six Agent Skills for goal writing, docs closeout and research

What Khazix Skills actually packages

The repository is a directory of six skills, not a library or a CLI. Each folder (leader, storage-analyzer, aihot, neat-freak, hv-analysis, khazix-writer) holds a SKILL.md, and the README describes every one of them as "a structured instruction set the Agent can load directly", following the Agent Skills open standard at agentskills.io. The README states that Claude Code, Codex, Qoder, Kimi Code, iFlow, CodeBuddy, Cursor and 40+ other agents that support the standard can install them.

The audience is narrow and that is a feature. These are skills the author says he runs in his own projects before open-sourcing them, so the set is shaped by one person's workflow: define a long-running goal, clean up project knowledge afterwards, scan a disk, pull an AI news digest, produce a research PDF, write in a specific voice. If your workflow looks nothing like that, most of the collection is dead weight. The parts worth evaluating on their own merits are leader and neat-freak, because both attack a failure mode that shows up in any long agent session.

leader: turning a vague request into a goal with a stop condition

The leader skill does one thing. It converts a fuzzy request into a written goal document that can be pasted into a goal mode (the README names Claude Code's /goal and Codex's goal mode) so the agent runs for hours without supervision. The README's argument is that the unit of human-AI collaboration has grown from a turn to a task to a goal, and that a goal which drifts for eight hours wastes more than a task that drifts for two minutes.

The mechanism is a seven-question checklist the skill applies before writing anything: purpose, completion state, evidence, anti-cheating rules, boundaries, trade-off ordering, and unknown territory. Two of those carry real weight. Evidence means every acceptance item has to paste actual command output, so the agent cannot declare success. Anti-cheating means the skill names the lazy paths in advance; the README's own example is that the cheapest way to make a test suite green is to delete the tests. Boundaries combine a whitelist with a hard round limit.

The README also describes a zeroth question: whether the map you are handing over was measured or merely heard. The skill therefore inspects the codebase itself before writing, on the stated grounds that a command written in a doc may not exist. The documented flow is: trigger, investigate, ask up to five questions only you can answer, then write the task book, about 12 minutes end to end. Output is plain Markdown, so agents without a goal mode can take the same paste. The README recommends a strong planning model for the goal and a strong execution model for the run, and lists the author's own pairing.

neat-freak and the three layers of project knowledge

neat-freak is the skill that runs after a task, triggered by /neat, and reconciles what the session changed against three separate layers: the root CLAUDE.md or AGENTS.md, which is read by the current agent; docs/ and README, which are read by colleagues; and the agent's own memory, which is read by future sessions. The README is explicit that these audiences do not overlap and must be handled separately.

It also audits rules as knowledge rather than just rewriting prose: whether CLAUDE.md and AGENTS.md are in sync, whether required files are missing, and whether paths referenced inside rules still exist. That last check matters more than it sounds. A rule that points at a deleted directory is worse than no rule, because the next session acts on a false premise.

The v3.0 behaviour documented in the README has two constraints worth quoting in spirit. Small projects with no git and no rule files get a lightweight path: README is aligned to the current code, a minimal AI rule file is created so the next session can recover context, and leftovers such as PLAN.md, debug scripts and xxx_old files are listed for confirmation. And nothing is deleted unilaterally: deletions are candidates only, machine-generated memory is read-only by default, and an instruction found inside a file is not treated as your authorization. That last clause is the interesting one. It means the skill treats file contents as data, not as commands, which is the correct posture for anything that writes to your repository.

Installing a skill by asking your agent to clone it

There is no package manager step. The README's install instruction is to tell the agent, in natural language, to install the skill by URL, and the agent clones it into the right directory itself. Replace the skill name in the path with the one you want, for example neat-freak, hv-analysis or khazix-writer.

帮我安装这个 skill：https://github.com/KKKKhazix/khazix-skills/tree/main/<skill-name>

The README does not document a manual clone path, a target directory, or a rollback procedure, so if you prefer to control placement yourself, the repository layout is the only guide: each skill is a top-level folder with a SKILL.md inside it, and you can copy that folder wherever your agent looks for skills.

The documented fallback for agents without skill support is to download the full SKILL.md and use it as a project rule file, or paste it into the conversation. The README claims the effect is the same, which is plausible for instruction-only skills like leader and khazix-writer, and less obviously true for storage-analyzer, which the README says opens a local browser report. A first real use for neat-freak is the closeout command, which the README lists alongside the trigger phrases.

/neat

After that you should see a change summary covering the three layers above, plus a candidate list for anything it wants to delete. Nothing is removed until you confirm.

Where the collection is the wrong tool

The README itself rules out two of the six for common cases, which is unusual candour. hv-analysis is documented as unsuitable for looking up a term, since ordinary conversation handles that; it is also documented as unsuitable for writing a WeChat article, which is khazix-writer's job. khazix-writer is documented as unsuitable if what you want is general good prose, because the skill takes positions: it refuses phrases like 赋能, 抓手 and 闭环, refuses 首先...其次, and refuses 在当今 AI 快速发展的时代. If your readers expect that register, the skill fights you.

The bigger constraint is storage-analyzer's platform coverage. The README states macOS is fully tested, while Windows code is ready with multi-drive support but first use should be treated with care. That is a documented testing gap, not a bug report, and it is the kind of thing to weigh before letting an agent propose deletions on a Windows machine. The read-only scan and the three-tier colour model (green deletable, yellow move-to-trash only, red open-folder only) are the safety story, but the README notes the local server binds to 127.0.0.1 on a random port with a token, and it does not document what happens if that server is left running or how the token is rotated.

The last limitation is structural. These are prompt-shaped skills, not verified software. The README describes what each skill should do; the SKILL.md is what the agent reads and improvises from. Nothing in the repository enforces the rules, so the quality of the outcome depends on the model executing it.

aihot versus wiring up an API yourself

aihot is the clearest contrast in the collection. It lets an agent pull the daily AI HOT digest and the full stream from aihot.virxact.com with no API key and no MCP server configuration. The documented capabilities include today's or a named date's digest, a curated item stream, current hottest events ranked by heat rather than recency, category pulls (model, product, industry, paper, tips), native 24-hour and 7-day windows, keyword, company and topic search, and syncing the current curated set locally so later requests only receive changes.

The alternative is building the same thing against a news API or an RSS pipeline. That gives you control over sources, ranking and storage, and it does not depend on one site staying up or keeping its agent endpoint open. The cost is that you own the deduplication, the heat ranking and the incremental sync logic, which is exactly the work aihot has already done. The trade-off is dependence on a single upstream: the README points to aihot.virxact.com and an agent onboarding page, and if that service changes its interface, the skill needs updating. For a personal daily digest that is a reasonable bet. For a pipeline feeding something you publish, the local-sync capability is the part that matters, because it means the agent stops re-fetching the whole set on every run.

Licence, maintenance and what upgrading costs you

The repository is MIT licensed, and the LICENSE file sits at the top level. MIT is permissive: you can copy, modify and redistribute the skills, including inside a commercial product, provided the copyright notice and permission notice are retained. That is a statement about the licence text, not legal advice; if you are folding these into a product, have someone check how you are attributing them. One practical detail for anyone vendoring a skill: because each skill is a folder with a SKILL.md, the natural upgrade is replacing that folder, which means any local edits you made are overwritten. There is no versioning scheme, no changelog and no release artefacts in the repository listing, so upgrades are manual and you should keep your modifications somewhere else.

On maintenance, the repository is not archived, and no last push date is published for it, so there is no basis for calling it actively maintained. Treat the current state as the state: six skills, one README in Chinese and one in English, no releases. The upgrade cost is mostly re-reading SKILL.md after you replace it, since behaviour lives in prose and prose can change without a version bump.

Editorial conclusion

Adopt khazix-skills if you already run an agent that loads SKILL.md files and you want the docs-and-memory closeout behaviour of neat-freak or the goal template from leader; both are plain Markdown and can be pasted into a chat if your agent has no skill loader. Skip it if you want a general-purpose writing assistant or a one-line definition lookup, since khazix-writer is deliberately opinionated and hv-analysis is documented as overkill for that. Before you commit, open neat-freak/SKILL.md and storage-analyzer/SKILL.md and read the deletion and scan rules yourself, because the README describes the safety model but the skill file is what your agent actually executes.

Frequently asked questions

Official sources

1. Official README
2. Project repository