20 of the most popular Claude skills, ranked
Skills are short instruction files that load when the work needs them. Most are public and free. There are now thousands, the quality varies wildly, and star counts tell you almost nothing about which ones are worth having. This is twenty of the best known, ranked S to D, with the reasoning for every placement.
A skill is a short set of instructions that sits on your machine until it is needed. When the work matches, it loads itself and changes how the job gets done.
Someone spends two years learning how to debug properly. They write that method down once. Everyone who installs it gets the careful version by default. That is the whole idea, and it is why skills are worth being deliberate about.
The problem is choosing. There are thousands of them now. Stars tell you how many people liked the idea of a thing. They do not tell you how much your work improves. Some of the most starred skills do very little. Some of the ones that matter most are barely known.
How we ranked them
Popularity decided which skills were eligible. It did not decide where they placed. A skill had to be well enough known to be worth naming. After that, two questions set the order.
- How much better is the output? A skill that rescues a poor result beats one that tidies a good result.
- How often do you do this? A big improvement to a yearly task is worth less than a steady improvement to a daily one.
When the two disagree, frequency wins. That single decision shapes the whole table, and it is where we are most likely to differ from other people ranking the same skills.
S tier
/systematic-debugging
A four phase process for finding the real cause. Left alone, an assistant guesses at a likely cause, changes something, and calls it fixed. You end up with three new bugs and the first one still there. This skill forces the slow version. Reproduce it. Isolate it. Understand it. Then fix it. Nearly everyone debugs every day, and this is where careless and careful are furthest apart.
/skill-creator
The skill that writes skills. It handles the parts people fumble, especially the description line that decides whether a skill ever loads at the right moment.
It ranks here for what it leads to. The deployment steps nobody wrote down. The review checklist that lives in one person's head. Each one becomes something the assistant applies every time. Most teams get more from two skills they wrote themselves than from twenty they installed.
A tier
/test-driven-development
Write the failing test, make it pass, then tidy up. Per use, this improves the output as much as anything on the list. A test written before the fix is the only real proof the fix worked. It sits below systematic-debugging on frequency alone. Far fewer people run it consistently than debug daily.
/code-review
Checks a change against the standards the project has written down, and against what you actually asked for. The second one matters more than people expect. A lot of AI assisted work is clean, tidy, and quietly solves a different problem.
/mcp-builder
Builds the connectors that let an assistant reach your real systems. Your calendar. Your CRM. Your database. This is what turns an assistant that writes about your business into one that works inside it. You do it rarely, which keeps it out of S. When you do, getting it wrong is expensive.
B tier
/writing-plans
Turns a vague job into a plan specific enough that building it is mechanical. A bad plan costs nothing to throw away. A bad build costs a week. On small jobs the planning step is pure overhead, which is why it sits here rather than higher.
/domain-modeling
Keeps the shared vocabulary straight. What a job means here. What an account means. What a run means. The payoff is slow, so it gets overlooked. It prevents the drift where one word means three things in three places, which nobody notices until something breaks.
/webapp-testing
Drives a real browser to check the change works in the app, not just that it compiled. This closes the most common gap in AI assisted work, which is being told something is fixed when nobody ever opened the page.
/using-git-worktrees
Stops parallel branches treading on each other. The moment you have two agents working at once, it earns its place. Before that it does nothing. Narrow, and very good inside that range.
C tier
Everything here is well made. It sits in C because the job is narrower or rarer, not because the work is weaker.
- /pdf. Pulls out text and tables, fills forms, merges and splits. Far better than improvising. Most people need it in bursts.
- /xlsx. The same for spreadsheets, formulas and formatting included. Excellent when the job lands, which is not often.
- /research. Digs into a question against real sources and writes the answer into the repository. The written record is the point. It stops the same question being answered three times.
- /frontend-design. Steers away from the generic AI look that makes everything resemble the same template. Useful, though taste is the part worth keeping for yourself.
- /playwright-skill. Solid browser automation, widely used. It overlaps with webapp-testing, which is what puts it a tier lower.
D tier
This is where we expect an argument. Two of the most starred skills anywhere are in it.
/get-shit-done
A spec driven system with tens of thousands of stars, and our least comfortable call. Structure is not the problem. The problem is that the ceremony costs the same on every job, including the ones that needed twenty minutes and a clear head. If your team ships large features over weeks, move it up several tiers and we will not argue.
/frontend-slides
Builds animated HTML presentations. It is popular because the output looks good in a screenshot. A presentation carries an argument, and animation does not improve a weak one. Most people also build slides rarely, so both of our questions point the same way.
/loki-mode
Runs dozens of agents as an autonomous startup system. It demos beautifully. Coordinating many agents is genuinely hard, and running a lot of them unsupervised mostly multiplies the places a wrong assumption can take hold.
/brand-guidelines
Applies Anthropic's own colours and typography. Useful if you work at Anthropic. For everyone else it is a good example to copy and an odd thing to install.
/algorithmic-art and /slack-gif-creator
Generative art, and GIFs sized for Slack. Both do what they say. Neither improves anything a business cares about, which is the axis this list is built on. Fun, and last.
Where to get them
Nine of the twenty are Anthropic's own. Add them with /plugin marketplace add anthropics/skills. Four come from obra's superpowers library, installed with /plugin install superpowers at claude-plugins-official. Three come from Matt Pocock's collection, installed with npx skills add mattpocock/skills. The last four are individual repositories from lackeyjb, gsd-build, zarazhangrui and asklokesh, each with instructions in its readme.
A ranking like this is a snapshot. Skills are new, the good ones are improving quickly, and anyone claiming a settled answer is overstating what the evidence supports. If you disagree with a placement, particularly the two at the bottom with all the stars, we would like to hear the argument.