Claude-Skills: a multi-domain prompt and tooling library for AI coding assistants
368 AI skills, 76 expert agents, and 859 stdlib Python tools for every team: engineering, PM, marketing, C-level, compliance, business ops, and research. Installs on Claude Code, Cursor, Codex, Gemini, Copilot, and 6 other AI assistants.
At a glance
- What is it?
- Borghei's Claude-Skills packages 368 skills, 76 expert agents and 859 stdlib-only Python tools across 20 domains, with a CLI that detects your assistant and drops files in the right directory. The breadth is the point, and the maintenance surface is the cost.
- Who is it for?
- Adopt Claude-Skills if you want one install to cover several teams at once, and if your workflows can tolerate prompt-and-checklist outputs rather than a hosted service. Skip it if you need a single deep, versioned toolchain for one discipline, since 368 skills spread over 20 domains means each one gets less scrutiny than a focused project would give it.
- Can I use it commercially?
- Check first. The repository uses a licence we do not classify automatically, so read its LICENSE file before any commercial use.
- Is it still maintained?
- Yes. The repository last received commits 34 days ago.
- What is it written in?
- Mainly HTML, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
Who Claude-Skills is aimed at, and the gap it fills
Most skill collections for AI assistants are engineering collections. They assume the person typing is a developer and the output is code. Claude-Skills is built on the opposite assumption: the README states it is "AI skills for every team, not just engineering," and the domain breakdown backs that up. Engineering is the largest single domain at 91 skills, but Project Management has 68, Marketing 39, C-Level 31 and Compliance 27. Legal, HR, Sales, Finance and Research Ops all have their own folders.
The problem it addresses is that a compliance officer, a product manager and a backend engineer all sit in the same company, all use the same assistant, and none of them want to write their own prompt scaffolding. The repository's pitch is that you install one thing and the right material lands in front of the right person. It also claims coverage of regulated industries that most open source skill libraries avoid: SOC 2, ISO 13485, MDR, FDA, EU AI Act, NIS2, DORA, HIPAA and GDPR, described as 18 compliance frameworks with checklists, gap analyses and audit-ready outputs. Whether those outputs survive a real auditor's review is not something the repository demonstrates, and the README does not claim they replace professional judgement.
How a skill, an agent and a bundle differ
Three separate abstractions sit on top of each other, and the README treats them as distinct products. A skill is a single unit of work. An agent is a persona that pulls several skills together: the README lists 8 cross-domain personas including startup-cto, growth-marketer, solo-founder, devops-engineer and finance-lead, activated with a slash command such as /persona startup-cto or by mentioning the name in conversation. On top of the cross-domain personas sit 68 single-domain cs-* agents such as cs-cto-advisor, cs-fintech-advisor and cs-scrum-master, which are narrower by design.
Bundles are the third layer and the one most likely to determine whether you get value on day one. bundles.json defines pre-configured sets by role: SaaS Founder Kit, DevOps Kit, Compliance Kit, Growth Kit, Product Kit, Data Kit, Security Kit, Finance Kit, PM Starter Pack and PM Career Track. Installing a bundle is meant to place the relevant subset of skills into your project rather than all 368. That matters, because dumping the full library into a repository is a lot of files to review and a lot of context for an assistant to sift through.
Installing it: the npx path, the desktop path, and the no-terminal path
The developer route is a single command: npx @borghei/claude-skills add senior-fullstack. The README states the CLI auto-detects Claude Code, Cursor, Codex, Gemini CLI, Copilot, Windsurf, Cline, Aider and Goose, and drops the skill into the right directory for whichever assistant it finds. It requires Node 18+ and no Python. That last detail is worth noting, because the 859 Python tools bundled elsewhere in the repository would otherwise imply a Python dependency at install time.
There are two other install paths for people who never open a terminal. In the Claude desktop app, you open Customize, go to Browse plugins, and enter borghei/Claude-Skills, which the README says makes all 368 skills available across Claude conversations. On the web, you browse the skill library site, click Try it now on any skill, and paste the prompt into Claude.ai, ChatGPT or Gemini, or click Save to my AI to install it as a Claude Project or Custom GPT. The README also points to docs/INSTALLATION.md for the Claude Code plugin, MCP server and manual install, and notes that 15 PM skills are callable as MCP tools.
One gap: the README contains an unfilled TODO for an install GIF, with a note that the recording recipe lives in .docs/images/README.md. The visual walkthrough does not exist yet.
What the 859 Python tools actually are
This is the part of the repository that separates it from a folder of markdown prompts. The README describes 859 Python tools that you can pipe into CI, MCP servers or your own scripts, and specifies that they use the standard library only, with no pip install step. In a corporate environment where adding a dependency requires a review, that constraint is the difference between usable and not.
The claim that they "actually run, not just prompt templates" is the kind of statement that needs verification against the source, which this review cannot do. What can be confirmed from the material is the shape of the packaging: the tools are counted separately from the skills, and the repository also ships a markdown-to-HTML publishing pipeline with what the README calls a CI-ready review gate. The Project Management domain is described as the deepest, with 68 skills, each shipping a worked example, an anti-pattern library and a deterministic scoring rubric, plus 5 chain pipelines and 3 live data adapters for Jira, Linear, Notion, Productboard and Confluence. A deterministic rubric is a meaningful design choice: it means the scoring is repeatable rather than model-judged, which matters if you want to compare results across runs.
The breadth problem: 20 domains, 368 skills, one release cadence
The release history shows how fast this moves. v4.2.0 added 3 domains, 41 agents and 21 skills. v4.7.0 rewrote Project Management into what the notes call a production toolkit with 54 skills and roughly 30k lines added. v4.8.0 added 43 skills across three tiers, a new research domain and 5 C-suite advisors. Each of those is a substantial content change inside a single minor version number.
That cadence has a direct cost for anyone pinning a version. If your team has adapted a PM skill to your own sprint format, a v4.7.0-scale rewrite of that domain will collide with your edits. The README does not describe a migration path, a deprecation policy or a compatibility guarantee between minor releases. The pragmatic reading is that this library is designed to be consumed at HEAD rather than pinned, which is fine for prompt content and less fine for anything wired into CI.
The second cost is depth per skill. With 368 skills across 20 domains, the average skill receives less attention than a project that ships 30. The README's own emphasis on Project Management as the deepest domain is an implicit acknowledgement that the other 19 are shallower. A legal team evaluating the 17 legal skills should read them before assuming the same quality bar applies as in the 68-skill PM set.
Licence and the Commons Clause
The badge in the README reads MIT + Commons Clause, while the repository metadata reports the licence as NOASSERTION, meaning GitHub's classifier could not map the LICENSE file to a standard identifier. Those two signals point at the same thing: this is not plain MIT.
The Commons Clause is an addition to a permissive licence that typically restricts selling the software. The practical consequence is that the permissive half of MIT (use, modify, redistribute) generally still applies, while commercial resale of the licensed work may not. The exact scope depends on the wording in the LICENSE file, which this review has not read and cannot summarise. If you plan to embed these skills in a product you sell, or to offer them as part of a paid service, read LICENSE yourself and get a lawyer's view rather than relying on the badge.
Alternatives, and where the approaches diverge
The obvious comparison is Anthropic's own Agent Skills format, since Claude-Skills is built on top of it. The difference is scope and opinion. Anthropic's documentation describes the mechanism (a folder with a SKILL.md and supporting files, loaded on demand). Claude-Skills supplies the content: 368 filled-in skills, 76 agents and a CLI that writes them into the correct directory for eleven different assistants. If you already know what skill you want, the format alone is enough. If you want a starting library that spans compliance, marketing and PM, the format alone gives you nothing.
A second comparison is the MCP server ecosystem. MCP servers expose tools to an assistant over a protocol; Claude-Skills ships 15 PM skills as MCP tools but is primarily a file-based library. The trade-off is state. An MCP server can hold a connection to Jira or Linear and return live data. A file-based skill is a document the model reads. The README mentions 3 live data adapters for Jira, Linear, Notion, Productboard and Confluence, which suggests the project is reaching toward the live-data model without abandoning the file-based one.
Orchestration, commands and the parts most people will ignore
Two features sit outside the main install flow and are easy to miss. The first is orchestration: 4 multi-agent patterns (sequential pipeline, parallel fan-out, supervisor delegation, consensus voting) defined in the project's Orchestration Protocol standard. The second is slash commands: 26 of them, including /tdd, /rice, /prd, /retro, /tech-debt, /security-scan, /sprint-plan and /changelog, with the full list available by running /README in Claude Code.
There are also three deep-dive packs that go further on one topic than the main library does: Playwright Pro with 8 sub-skills, Self-Improving Agent with 6 sub-skills for agents that evaluate their own outputs, and AgentHub with 7 sub-skills for multi-agent orchestration. These are the parts of the repository where a single subject gets more than a single file, and they are probably the best signal of what the project looks like when it is not spreading itself thin. Consensus voting in particular implies multiple model calls per decision, which has a cost profile the README does not quantify.
Editorial conclusion
Adopt Claude-Skills if you want one install to cover several teams at once, and if your workflows can tolerate prompt-and-checklist outputs rather than a hosted service. Skip it if you need a single deep, versioned toolchain for one discipline, since 368 skills spread over 20 domains means each one gets less scrutiny than a focused project would give it. Before committing, read the LICENSE file to see what the Commons Clause restricts, check the changelog to see how often domain content is rewritten between minor versions, and run npx @borghei/claude-skills add senior-fullstack on one machine to confirm which directory the CLI targets for your assistant.
Community notes