Claude Skills for Developers: 8 Worth Installing, by Job

The best Claude skills for developers aren't the ones that make Claude "smarter at coding." They're the ones that take a job you keep re-explaining — how to commit, how to check the page actually works, what this repo's build command is — and write it down once, so the agent does it the same way every session.

This roundup is organised by that job, not by who published the skill. Eight skills, each from a public GitHub repository, three of them from Anthropic's own collection:

• Commit hygiene: Git Workflow & Versioning (addyosmani/agent-skills)
• Database work: SQL Queries (phuryn/pm-skills)
• Testing a web app: webapp-testing (anthropics/skills)
• Onboarding an agent to a repo: AGENTS.md generator (microsoft/skills)
• Prototyping: Prototype (mattpocock/skills)
• Architecture diagrams: Archify (tt-a1i/archify)
• Writing your own skills: skill-creator and mcp-builder (anthropics/skills)

Every description below comes from the skill's own files. Nothing here is a star count or a popularity ranking.

Quick comparison

If you only install one, pick the one that matches the correction you typed most often last week.

Commit hygiene: Git Workflow & Versioning

An agent that writes 800 lines and commits them as update files has given you something you can't review and can't cleanly revert. The Git Workflow & Versioning skill in Addy Osmani's agent-skills repository makes it work the other way.

What it enforces:

The part developers running parallel agents get the most from is the section on worktrees — each task in its own worktree, so a failed experiment gets deleted with nothing lost. We cover the full skill in the Git Workflow & Versioning skill guide.

Database work: SQL Queries

The usual failure with AI-written SQL isn't syntax. It's a confident query against a column that doesn't exist.

The SQL Queries skill from Pawel Huryn's pm-skills repository starts from your schema instead: table names, columns, types, keys, and relationships. It then confirms your dialect — BigQuery, PostgreSQL, MySQL, Snowflake, or SQL Server — asks about filters and aggregations, and returns a commented query with a plain-English explanation and performance notes.

Two limits matter. It never connects to your database, and it only sees the schema you hand it. You still run the query yourself, ideally with a row limit or against a copy first.

Testing a web app: webapp-testing

"Does it actually work?" usually gets answered by someone opening a browser and clicking around. Anthropic's webapp-testing skill, in the anthropics/skills repository, hands that job to the agent and makes it leave evidence.

Its core pattern is reconnaissance, then action. The agent loads the page, takes a screenshot or reads the rendered DOM, works out real selectors from what's on screen, and only then clicks or types. It waits for network activity to settle before inspecting a dynamic app, which is the difference between a check that fails at random and one that doesn't. A helper script starts your server — or a backend and frontend together — runs the test, and shuts everything down.

Upstream, the agent writes native Python scripts for Playwright, so you need Python, the Playwright library, and its browser binaries installed first. Taku's marketplace carries an adaptation that swaps that stack for browser tools Taku manages; the Web App Testing skill page explains how the two differ.

It doesn't replace an end-to-end suite in CI. It's the in-session check before you merge.

Onboarding an agent to a repo: the AGENTS.md generator

A coding agent that keeps guessing the wrong test command doesn't need a better model. It needs the command written down.

AGENTS.md is the file coding agents read for that: build commands, test commands, structure, conventions, limits. Microsoft's skills repository includes a skill in its deep-wiki plugin that writes these files folder by folder.

The first row is why it's safe to point at a real codebase: nothing you've tuned by hand gets overwritten. The flip side is that it generates rather than audits, so it won't improve files you already have. More in the AGENTS.md generator guide.

Prototyping: Prototype

The Prototype skill in Matt Pocock's skills repository builds throwaway code that answers one design question, then gets out of the way.

It first decides which question you're asking, because the two need different artifacts:

• "Does this state model hold up?" One shareable HTML file with free-play buttons and guided walkthroughs that push the logic through its awkward cases.
• "What should this look like?" Several radically different UI variations on one route, switchable through a URL parameter.

Six rules keep the output honest: labelled as throwaway, trivial to run, no persistence, no polish, full state visible, and the decision written down. The prototype itself ends up on a throwaway branch with a pointer from the issue.

One install note: it's a folder with a main SKILL.md and two branch guides. Copy only the main file and the skill loads, then can't find its guides.

Architecture diagrams: Archify

Ask most agents for an architecture diagram and you get either a flat sketch or hand-placed boxes with arrows running through them. Archify takes a different route: the agent writes a typed JSON description of the system, and a renderer compiles it into a single interactive HTML file.

It supports five diagram types — architecture, workflow, sequence, data flow, and lifecycle — and its validation rejects edges crossing unrelated boxes, ambiguous shared routes, and labels covering lines. The output has themes, pan and zoom, search, and relationship tracing. It can also start from an existing Mermaid diagram.

Upstream it's a Node.js project that can analyse an open repository. The version in Taku's marketplace is a reduced-permission adaptation that works from your written description and doesn't inspect repositories or start servers.

Either way, Archify checks that a diagram is well-formed. It can't check that your system really works the way you described it.

Anthropic's own: skill-creator and mcp-builder

Two more from the anthropics/skills repository are worth knowing, both in its example-skills plugin.

skill-creator writes and improves skills. Its loop: capture what the skill should do and when it should trigger, draft it, run test prompts with the skill loaded, grade the results, rewrite, repeat. It also has a description-optimisation step that builds a set of should-trigger and should-not-trigger queries — the near-misses matter most — because the description is what decides whether Claude loads a skill at all. Once you notice you're re-explaining the same process, this is the one that turns it into a skill.

mcp-builder guides building an MCP server in four phases: research and planning, implementation, review and testing, then writing evaluations. It covers Python (FastMCP) and TypeScript, recommending TypeScript by default. Use it when a skill isn't enough because Claude needs a new connection, not new instructions.

Anthropic labels the repository as demonstration and educational material and notes that behaviour inside Claude may differ from these implementations. Treat them as strong patterns, not a supported product.

How to install Claude skills for developers

In Claude Code, a skill is a folder with a SKILL.md. Put it in ~/.claude/skills/ to use it in every project, or in .claude/skills/ inside a repository to share it with everyone who works there. Anthropic's collection can also be added as a plugin marketplace with /plugin marketplace add anthropics/skills.

Commit hygiene and AGENTS.md belong at project level, because the whole team benefits. SQL and diagrams can live at personal level.

Don't install all eight at once. Every skill's description competes when Claude decides what to load. Our best Claude skills roundup covers the non-developer side — documents, design, brand — if you need those too.

Running a developer skill without the install step

Every skill above has the same hidden cost: finding the folder in someone's repository, copying the whole folder, picking a scope, and for two of them standing up Python or Node first. For one skill that's minutes. For a teammate who's new to coding agents, it's usually where they stop.

Taku is an AI-native desktop workspace built around that gap: you mirror a skill or workflow someone already got working, run it, remix it, and keep it. Its marketplace carries versions of several skills here, including the adapted Web App Testing and Archify. It doesn't remove your repository, your agent, or your judgement about what to merge — only the setup. Taku is in Beta, and the Mac app is available now.

FAQ

What are the best Claude skills for developers?

Pick by job. Git Workflow & Versioning for commits, SQL Queries for database work, webapp-testing for browser checks, the AGENTS.md generator for onboarding an agent to a repo, Prototype for design questions, and Archify for diagrams. Add skill-creator once you want your own.

Do these skills work outside Claude Code?

Most are written as general agent skills in the SKILL.md format, and Archify publishes install commands for several agents. Check each repository's README for the agents it supports.

Are these skills free?

The upstream skills are published in public GitHub repositories; check each repository's licence before redistributing. You still need your own coding agent to run them.

Do any of them touch production?

SQL Queries never connects to a database, and the AGENTS.md generator won't overwrite existing files. webapp-testing runs against a local app you point it at. You stay in control of what gets run and merged.

What's the difference between skill-creator and mcp-builder?

skill-creator writes skills, which carry instructions. mcp-builder builds MCP servers, which give Claude new connections to APIs and services.