Skip to content
Mia Automation
NLEN Sign in Put Mia to workGet started

The Best Claude Skills: We Tested Ten Against Claude Code Without Them

Of ten popular skills from Anthropic's skills repository and the Superpowers collection, three made a real difference in our test: test-driven-development, frontend-design, and internal-comms if you like its format. We gave Claude Code the same task twice per skill, once with the skill installed and once without, and scored what the skill added. Four skills produced the same result as plain Claude Code at 1.4 to 2.3 times the cost. Three we couldn't score, and the reasons are below too.

Claude in practiceUpdated 7 min read

Tested by us on 11 November 2026.

The scores

Animated scoreboard: test-driven-development 5 kept, frontend-design 4 kept, internal-comms 3 kept, systematic-debugging 2 and brainstorming 2 dropped, pdf 1 and docx 1 dropped, and no score for claude-api, verification-before-completion and skill-creator, with the cost of each run with the skill relative to without
What each skill added over the same task without it, and how much more the run with the skill cost.
Claude Code 2.1.321 on Ubuntu, November 11, 2026, one run with and one without per skill. Cost at API rates as reported by Claude Code. Score 1 to 5 for what the skill added.
SkillSourceTest taskWithout the skillWith the skillScore
test-driven-developmentSuperpowersAdd a Dutch postcode validatorCode and a quick manual check, no test file ($0.038)22 tests written first and seen failing, then the code; all pass ($0.084)5, kept
frontend-designAnthropicOne-page site for a painting companyTidy page, but colour names didn't match the swatches ($0.108)Bolder layout with a working colour picker, no mismatches ($0.116)4, kept
internal-commsAnthropicWeekly status update from given factsClear update with headings ($0.028)Short Progress, Plans, Problems format ($0.041)3, kept if you want that format
systematic-debuggingSuperpowersFix a bug where the second invoice total is too highRoot cause fixed: a shared default list ($0.041)Same fix ($0.063)2, dropped
brainstormingSuperpowers"Automate payment reminders. Build it."Asked four questions first ($0.033)Asked one open question at a time ($0.054)2, dropped
pdfAnthropicMerge two PDFs, add "Page X of Y"Correct, 2 pages numbered ($0.041)Same result ($0.056)1, dropped
docxAnthropicQuote letter with heading, table and footerAll three present ($0.044)Same result, 22 s instead of 9 s ($0.099)1, dropped
claude-apiAnthropicScript that streams a Claude API answerAlready built into Claude Code: it loaded anywaySameNo score
verification-before-completionSuperpowersSmall change in a project with two failing testsNo test run ($0.045)Skill never loaded, though this run did test ($0.165)No score
skill-creatorAnthropicCreate a skill for meeting notesWrote a valid skill ($0.040)Write to the skills folder was blocked in our headless run ($0.096)No score

How we tested and scored

Each skill got a task it was written for, in a fresh, empty project. For the run with the skill we copied its folder into .claude/skills/; for the run without, the folder was absent. Where we could, we checked the outcome with code: page numbers read back from the PDF, the heading, table and footer read from the Word file, the bug re-run, the tests counted. The score is what the skill added over the run without it: 5 is a clear gain, 1 is the same result at a higher price.

Keep one thing in mind when you read the table: Claude Code 2.1.321 already ships with built-in skills, including debug, verify, code-review, design and claude-api. So the "without" runs weren't bare. A skill has to beat Claude Code as it comes, which is the comparison that matters when you decide what to install.

The three we kept

test-driven-development changed behavior more than any other skill. Without it, Claude Code wrote the function and checked a few cases by hand. With it, Claude Code wrote 22 tests first, ran them against an empty stub to see them fail, then wrote the code until they passed. That cost 2.2 times as much and left a test file behind, which is what you're paying for. One thing to know: to run the tests it fetched pytest on the fly with uv, so it downloaded a package without being asked.

Two website screenshots side by side for the painting company: left a cream page with a headline and paint swatches whose names don't match their colours; right a blue page with a very large Schilder Bos headline and a row of named colour swatches (opens at full size)
frontend-design, without and with the skill. Both pages are in Dutch because the company is Dutch. Screenshots taken in Chrome on a Windows PC.

frontend-design gave a more distinctive page for almost the same cost. The page without it looked good at a glance but labeled a dark green swatch "Gebroken wit" (off-white) and a beige one "Mosgroen" (moss green). The version with the skill picked one idea, choosing a wall colour, and carried it through. Taste is subjective; the mislabeled swatches aren't.

internal-comms turns facts into a short Progress, Plans, Problems update. It's built around how Anthropic writes internally, so keep it if that format suits your team and skip it if you already have a template.

Why the others dropped out

  • pdf and docx produced the same files as plain Claude Code. They're the skills behind Claude's own document features and are much longer than our tasks needed. On a complex form or a tracked-changes edit they may earn their keep; on everyday documents they cost more for the same result.
  • systematic-debugging found the same root cause plain Claude Code found: a shared default list in Python. A bug with a misleading trail might show a difference; ours didn't.
  • brainstorming asked better-paced questions, one at a time instead of four at once. It also ships a small web server for a visual companion, which we blocked for this test. That's a lot of machinery for a nicer conversation.
  • claude-api is already built into Claude Code, so installing it adds nothing.
  • verification-before-completion and skill-creator we couldn't score: the first was never loaded in our run, and the second hit a permission block writing to the skills folder.

Not tested: xlsx and pptx (their Python libraries weren't on our test machine), webapp-testing and web-artifacts-builder (they drive a browser or a build step we don't run on servers), and brand-guidelines, which applies Anthropic's own brand colors.

Install only the skill you want

Both collections install as plugins, but a plugin brings every skill in it, and Superpowers also brings hooks. If you want one skill, copy its folder into .claude/skills/ in your project or ~/.claude/skills/ for all projects, and start a new session.

bash
git clone --depth 1 https://github.com/obra/superpowers.git /tmp/superpowers
git clone --depth 1 https://github.com/anthropics/skills.git /tmp/anthropic-skills
mkdir -p .claude/skills
cp -r /tmp/superpowers/skills/test-driven-development .claude/skills/
cp -r /tmp/anthropic-skills/skills/frontend-design .claude/skills/

Copy the two skills we kept from their repositories. Install commands for the full plugins are in each repository's README.

Check a third-party skill before you install it

  1. Read the SKILL.md. It's instructions Claude Code will follow as if you wrote them. Look for steps that send data somewhere or run commands you didn't expect.
  2. List the other files. Scripts in a skill run with your account's rights when Claude Code calls them. The brainstorming skill, for example, includes a Node.js web server.
  3. Check what it fetches. In our test the TDD skill led Claude Code to download pytest. Harmless here, but on a work machine you may want to know first.
  4. Check the license. Anthropic's README says the skills in its repository are open source under Apache 2.0, including the document skills. Superpowers is MIT. Other collections vary, so look before you copy.

Questions

What are Claude skills and how do I install them?

A skill is a folder with a SKILL.md that Claude Code loads when a task matches its description. Copy the folder into .claude/skills/ or ~/.claude/skills/, or install a whole collection as a plugin from its marketplace.

Which Claude skills are worth it?

In our test: test-driven-development and frontend-design clearly, internal-comms if you like its format. The rest matched plain Claude Code at a higher cost, or couldn't be scored.

Are third-party skills safe?

They're instructions and scripts that run as you. Read the SKILL.md and the scripts, check what they download, and install single skills rather than whole plugins when you only need one.

Do skills work in the Claude app and in Claude Code?

Anthropic's skills README links to guides for using skills in the Claude apps as well as in Claude Code. We only tested in Claude Code. We have no paid relationship with Anthropic or the Superpowers author.

Sources