ClaudeSkills: 13 Agent Skills That Treat Prompts as Software

The Problem: Prompts Die, Skills Persist

A prompt works once and then vanishes with the conversation. Each new session starts from zero. ClaudeSkills addresses this by packaging methods as Agent Skills: directories of files that include steps, templates, scripts, examples, and acceptance checks. The README makes the distinction explicit: a prompt lets a model do something once, then dies; a skill keeps the method around so it can be reused, inspected, and improved. This is for engineers and product teams who find themselves repeating the same elaborate instructions to an AI agent, whether that is researching a topic, drafting a product spec, or building a slide deck. Instead of retyping the process, you install a skill and reference it by name.

The Architecture: SKILL.md and a Directory of Artifacts

Each skill is a directory with a SKILL.md file at its root. The format is open, meaning any client that supports Agent Skills can read it. The skill contains the complete process: the steps the agent should follow, the templates it fills in, the scripts it runs, and the acceptance checks it must pass. The repository layout shows this clearly. For example, deck-studio keeps its generators, rendered pages, rubrics, and review notes inside the skill directory. The skill is not a single prompt; it is a small software project. The install script copies the whole directory to ~/.agents/skills/ by default, and the directory name becomes the skill name. This design means you can inspect exactly what the agent will do before you let it run, and you can modify the files to change the behavior.

Getting Started: Install, Name, and Run

The quick start is straightforward. Clone the repository with git clone --depth 1 https://github.com/staruhub/ClaudeSkills.git && cd ClaudeSkills, then run python3 scripts/install_skill.py deck-studio. After that, you tell your agent: "Use deck-studio to turn this quarterly review into an 8-slide consulting deck." The skill installs to ~/.agents/skills/ by default, which most clients scan. There are other options: --project installs to the current project only, --client claude-code targets Claude Code's directory, and manual copy is possible with cp -r skills/Geek-skills-deep-research ~/.agents/skills/deep-research. To update, run git pull and then reinstall with --force. The README warns that installed skills are copies, so you must reinstall after pulling changes. Uninstalling is just rm -rf ~/.agents/skills/deck-studio.

The 13 Skills: From Deep Research to C-Drive Cleanup

The curated set has four flagships. deep-research scopes the question, gathers sources, registers each one, and checks citations, then hands back a report with references and limits. product-manager reads the repo first, asks one question per round, and turns a vague idea into a product document, with no code before the doc is approved. deck-studio builds slide decks with outlines, per-page briefs, registered layouts, and visual checks. wechat-article-writer produces copy, image prompts, and layout HTML for WeChat articles, but never publishes for you. The nine professional skills cover pair-programming with structured self-review, security audits, solution architecture, Three.js performance tuning, PDF parsing with MinerU, AI sales pitches, and two named development methods, keqian-method and xuefeng-method. There is also c-drive-cleaner for Windows cleanup, which is dry-run by default. The range is wide, which is both a strength and a weakness: you get a lot of ground, but each skill is only as deep as its own files.

Verification: Tests That Prove Structure, Not Quality

The project treats skills like software, and it provides scripts to check them. python3 scripts/validate.py verifies the directory structure of the 13 curated skills. python3 scripts/run_routing_evals.py runs 91 routing cases across 10 skills. python3 tests/task_b/run_contract_tests.py runs 17 contract cases covering the success, resume, and failure paths of all four flagships, and deck-studio additionally runs real Chrome rendering and PPTX assembly. The README is honest about what these tests do not prove: output quality on a different model, real image generation, or actual WeChat publishing. The self-test scores in the repo, like the Constructivist example scoring 7.1/10 on its own rubric, are data points, not guarantees. You can rerun those tests yourself because the process and data are in the repo. This is a genuine attempt at verification, but it is verification of structure and fixed cases, not of real-world performance.

Limitations and Wrong Use Cases

The biggest limitation is that skills are copies, not live references. If you modify a skill after installing it, a git pull will not update it unless you reinstall with --force. Also, there is no automatic trigger: you must name the skill in your request, and the trigger mechanism varies by client. Some clients match natural language, others want slash commands. The README notes that if a skill is installed but not found, you need to check whether your client scans ~/.agents/skills/. This makes the system less plug-and-play than it first appears. Another limitation is that the verification checks are narrow. The contract tests cover only 17 fixed cases, and the routing evals cover 91 cases across 10 skills, but that does not cover every real-world input. The README explicitly says the checks do not prove routing accuracy when a real model runs. So if you need a skill to work perfectly in production with varied inputs, you will need to test it yourself. The experimental skills in lab/ skip the checks entirely, so treat them as prototypes.

Alternatives and Maintenance Costs

The obvious alternative is writing your own prompts or building your own skill directories from scratch. That gives you full control but no shared structure or validation. Another alternative is using a different skill collection, though the README does not name any. The difference here is the emphasis on verification: ClaudeSkills ships with scripts that check structure, routing, and contract cases, which most ad-hoc prompt collections do not. The maintenance cost is real. The repository is actively maintained, with release 1.0.1 pushed on 2026-08-12, and the last push on the same day. Updates require a git pull and a reinstall of any skill you use. The license is MIT, which means you can use, modify, and redistribute the skills freely, with attribution. There is a SECURITY.md file, but its contents are not detailed in the README. The project has a website and a release notes page, which suggests ongoing attention. You should budget time to rerun the validation scripts after any update, because the repo's own tests are the only guarantee you get.

Editorial conclusion

Adopt ClaudeSkills if you want a set of pre-built, testable Agent Skills for common tasks like deep research, deck creation, product documentation, and WeChat articles, and you are willing to run the provided validation scripts before trusting them. Do not adopt it if you need production-ready output for every skill, because the self-tests cover only a fixed set of cases and do not guarantee real-world quality. Before using any skill, verify that your client scans ~/.agents/skills/, run python3 scripts/validate.py and the contract tests, and check the verification/2026-07-31/README.md for the exact limits of what each test proves.

Official sources

1. Official README
2. Project repository
3. Release notes