Six skills solve six different jobs
The repository contains 6 modules for Agent Skills clients. leader writes a bounded goal brief, while neat-freak reconciles code changes with docs, rules, and agent memory. The other four cover storage, Chinese AI news, research PDFs, and Chinese long-form writing. That range suits the author's workflow, but each directory needs its own permission and quality decision.
The README tells you to give an agent one directory URL, or copy its SKILL.md into a client without native skill support. There is no collection account or central service. At commit b81ad3b, the host agent still supplies the shell, browser, filesystem access, web tools, and judgment. The instructions describe guardrails; your agent and environment must enforce them.
Leader and neat-freak are the safest starting points
leader and neat-freak are the 2 modules with the narrowest local consequences. leader inspects the project, defines completion, names forbidden shortcuts, and writes a runnable brief. Its seven questions cover proof, boundaries, tradeoffs, and unknowns. One open issue reports repeated runs that asked zero questions and selected defaults when the user expected a decision checkpoint.
neat-freak addresses stale project knowledge. The skill separates README files, agent rules, and saved memory, then requires authorization for writes or deletion. It has a light path for small projects and a fuller audit. Our 3-CPU, 8 GB sandbox did not execute either workflow, so these design claims come from the repository rather than an observed run.
What happened when we ran it
Our sandbox did not run Khazix Skills. We checked commit b81ad3b on 2026-09-25 in a fresh, unprivileged Debian container with 3 CPUs, 8 GB of RAM, and no secrets. Our method checked the checkout for a supported ecosystem and Dockerfile before any install attempt. The runner found neither.
We measured no install, build, or test result because there was no honest command to run. Five directories are mainly instructions and references, while some add scripts or outside calls. A top-level test would say little about whether leader asks the right question or hv-analysis cites a disputed fact correctly. Buyers also lack one repeatable health check for all six skills.
Windows cleanup is documented but unverified on Windows
The README says the Windows cleanup path did not complete real-hardware testing. storage-analyzer scans disk use, produces an HTML report, and can expose buttons for trashing or deleting approved paths. Its local server binds to 127.0.0.1, uses a random token, validates real paths, keeps targets inside the home directory, and requires browser confirmation.
The skill says its macOS scan, report, and deletion path were tested. Its Windows implementation had not been tested on real Windows hardware at commit b81ad3b, and issue 49 reports slow scanning. Windows also needs Python 3. Inspect every proposed path and test trash behavior with disposable data before trusting a whole-drive cleanup session.
Deep research still needs source discipline
Issue 67 documents 4 cases and inconsistent source labeling. hv-analysis researches a subject's history and current competitors, then generates a styled PDF through a bundled Python script using WeasyPrint and Markdown. The workflow depends on web search quality, source selection, package installation, and the model's handling of disagreement. A polished PDF can hide uneven evidence.
The reporter found systematic confidence labels in only one report and documented a wrong date taken from a real page with changed metadata. The proposed fix requires source classification and independent confirmation for important facts. Until that lands, treat the PDF as a research draft. Our 2026-09-25 lab result includes no report-quality measurement because the repository was not run.
Two skills remain Chinese by design
Two of the 6 modules are Chinese by design. khazix-writer encodes the author's Chinese WeChat style, examples, banned phrases, and self-checks; the English README says it is unsuitable for English output. AIHOT serves Chinese-curated news through an anonymous API, with the skill pointing to aihot.news and retaining the Virxact domain as a compatibility endpoint.
Translating either skill would change the examples, prompts, editorial choices, and writing rules. English writers can study how the process is encoded, but the bundled output is the wrong product for them. None of this was execution-tested in our 3-CPU sandbox.
Active commits outweigh the lack of a collection release
GitHub showed 20,940 stars, 2,215 forks, and 49 combined issues and pull requests when fetched on 2026-09-26. The last push was 2026-09-25, and recent issues discuss research confidence, Windows scan speed, and leader's question behavior. GitHub's latest-release endpoint returned no collection-wide release, although older neat-freak tags exist.
Use the commit history and each skill's instructions as the version boundary. You can install one directory without adopting the other five. Choose a workflow, read its permissions, and verify its result. Khazix Skills works best as six authored operating procedures. Its 20,940 stars do not supply a shared test suite.