Agent skills by task
Categories tell you what a skill is about. This tells you what it is for. Each page below covers one job people actually hand to an agent — what a good skill for it does, what to watch for, and the skills in the registry that do it.
Engineering
- Skills for code review A code review skill gives your agent a fixed checklist instead of whatever it happens to notice: correctness first, then security, then style, with a severity it has to justify.
- Skills for writing tests Test-writing skills are mostly about restraint: what to assert, what to mock, and when a test is testing the framework rather than your code.
- Skills for debugging Debugging skills exist because agents, like people, want to fix the first plausible cause and move on.
- Skills for refactoring A refactoring skill's job is to keep behaviour fixed while structure changes, and to keep the diff small enough to review.
- Skills for security review Security skills teach an agent to read code the way an attacker does: follow untrusted input to where it is executed, rendered, or concatenated into a query.
- Skills for API integration Integration skills save you from re-explaining an API every session: endpoints, auth shape, pagination, rate limits, and the three fields that are documented wrong.
- Skills for database and SQL work SQL skills are worth installing for the parts an agent reliably gets wrong under time pressure: index-aware query shapes, migration ordering, and the difference between a query that is correct and one that will still be correct at ten million rows.
- Skills for git and version control Git skills mostly encode a team's conventions — commit message format, branch naming, when to squash — so the agent stops producing 'fix stuff' and stops rebasing things it should not.
- Skills for frontend and UI work Frontend skills are how you stop an agent from inventing a fourth button style.
- Skills for performance optimization Performance skills are valuable mainly for the discipline of measuring first.
- Skills for prompt engineering Prompt skills are meta-skills: they teach the agent how to write and evaluate instructions, including its own.
- Skills for building MCP servers MCP skills cover the part of building a server that is easy to get subtly wrong: tool naming and descriptions that an agent can actually choose between, argument schemas that fail loudly, and transports.
- Skills for working with legacy code Legacy skills are for the situation where the code works, nobody knows why, and the tests do not exist.
- Skills for mobile app development Mobile skills carry platform conventions an agent will otherwise approximate from web habits: lifecycle, navigation patterns, what belongs on a background thread, and the store rules that reject a build.
Writing & docs
- Skills for writing documentation Documentation skills fix the two failure modes agents fall into unprompted: describing what the code obviously does, and inventing behaviour it does not have.
- Skills for PDF and document handling Document skills cover the work that looks trivial and is not: extracting text from a PDF that has columns, filling a form, producing a .docx that opens correctly in Word.
Data
- Skills for spreadsheets and Excel Spreadsheet skills are mostly about fidelity: preserving formulas rather than flattening them to values, keeping number formats, not silently coercing a leading zero away.
- Skills for data analysis Analysis skills are worth having for the steps people skip: checking what is missing, saying whether the sample supports the claim, and refusing to report a difference as meaningful when it is not.
- Skills for data visualization Visualization skills carry the judgments that make a chart readable: which form fits the comparison, how many colours a category scale can hold, when a number is better than a plot.
- Skills for web scraping and research Scraping and research skills give the agent a method for gathering: what to collect, how to cite it, and when a source is not good enough to use.
Design & product
- Skills for design and mockups Design skills give an agent taste it does not have by default — a type scale, spacing rhythm, a restrained palette, and rules about when to stop.
- Skills for accessibility audits Accessibility skills push an agent past the automated third — contrast and alt text — into the part that needs judgment: focus order, what a screen reader actually announces, whether a custom control has the semantics of the native one it replaced.
Operations
- Skills for CI/CD and deployment Deployment skills turn a runbook you half-remember into something the agent follows the same way at 2am.
- Skills for Docker and containers Container skills are about the difference between an image that works and one that is small, cached, and reproducible: layer ordering, multi-stage builds, what belongs in .dockerignore.
- Skills for shell scripting and automation Shell skills are about writing scripts that fail correctly: quoting that survives spaces in filenames, errexit and pipefail, and a dry-run mode for anything destructive.
Business
- Skills for SEO and content SEO skills are worth judging by how much they refuse to do.
- Skills for email and communication Communication skills are about voice and triage more than grammar: what deserves a reply, how long it should be, and what your team's register actually sounds like.
- Skills for project and task management Planning skills turn a vague request into a document with a shape: scope, sequence, what is explicitly out.
Frequently asked questions
What can AI agent skills do?
Anything you can write down as a repeatable method. In practice the common ones cover engineering work (code review, tests, debugging, refactoring, deployment), document work (PDFs, Word files, spreadsheets, documentation), data work (SQL, analysis, charts) and business work (email, planning, research). A skill is instructions, not a program, so the limit is what your agent can already reach.
How do I choose between two skills for the same task?
Read both SKILL.md files — every skill page shows the whole thing. Prefer the one that is specific about when NOT to apply, names the tools or frameworks it assumes, and declares the narrowest set of capabilities. Install counts and ratings help, but a skill that matches your stack beats a more popular one that does not.
Can I install more than one skill for the same task?
Yes, and it is common. Agents load skills on demand based on their descriptions, so overlapping skills mostly cost disk space rather than context. If two skills genuinely conflict on method, the clearer description usually wins the match — which is an argument for keeping only the one you trust.