frontend-design
49 / 100
anthropics/skills
Rubric met 95.9% → 91.7%
- No clear effect
How we evaluate
We rate skills by running them
Scores on cccskills are not based on reading SKILL.md. We give the same task to Claude Code with and without the skill and compare the results.
Reading alone did not tell skills apart
We used to have AI read only the SKILL.md and grade four points A, B, or C. When we tallied the results in October 2026, the scores barely separated anything.
68%
scored 90 or higher
Out of 2,839 rated skills. 694 of them scored a full 100.
98%
got an A for clarity
A well-written file does not show whether the skill works.
84%
got an A for concrete steps
Having steps written down is not the same as better results.
How we test
We test one popular skill a day, in the steps below. Claude is used only to write tasks, run them, and grade them. Totals and storage are handled by plain scripts.
- 01
Write the tasks
From the SKILL.md we write requests the skill should help with, starting files for a test project, and a rubric, and fix them before any run. The request never names the skill.
- 02
Run in isolation
Each run gets a fresh container. It can only reach the Anthropic API, so it cannot touch other sites or anything on cccskills.
- 03
Run with and without
The same task goes to Claude Code without the skill and with it. Which one runs first alternates.
- 04
Measure time and tokens
We record elapsed time, tokens, cost at API prices, number of turns, and whether the skill was actually loaded.
- 05
Grade blind
A separate AI, not told which run used the skill, scores each rubric item 1, 0.5, or 0. It checks the files produced, not the run's own description.
- 06
Publish the outputs
We publish every final message and every file created or changed, so anyone can check the evidence behind a score.
The run score (100 points)
The sum of five parts. We grade strictly: a skill that changes nothing ends up around 55.
| Quality | 45 | Difference in rubric points met. No difference is 15, +50 points is full marks, −25 points is 0. |
|---|---|---|
| Speed | 15 | Change in elapsed time. No change is 7.5, half the time is full marks, 1.5× the time is 0. |
| Tokens | 15 | Change in tokens, scored the same way as speed. |
| Activation | 15 | Share of runs with the skill installed in which Claude actually loaded it. |
| Stability | 10 | Share of runs that finished (5) and share of pairs the skill did not lose (5). |
Example: same quality, time, and tokens, always loaded, never stopped = 15 + 7.5 + 7.5 + 15 + 10 = 55
What the skill changed
Alongside the score, labels show the kind of change. A skill can have more than one.
Good changes
- Better output
- Rubric points met went up by 15 or more
- Faster
- Elapsed time fell by 20% or more without lower quality
- Fewer tokens
- Tokens fell by 20% or more without lower quality
Changes to watch
- Worse output
- Rubric points met went down by 10 or more
- Slower / more tokens
- Elapsed time or tokens grew to 1.5× or more
- Rarely loaded
- Not loaded in more than half of the runs where it was installed
- Runs stopped
- One run in four or more did not finish
What we do to keep it fair
These conditions keep the score close to what you would see in real use. We list the limits too.
Blind grading
The grader is not told whether A or B used the skill, and the order is random each time.
Same conditions
Same task, same starting files, same model (Claude Sonnet). The only difference is the skill.
Stars are ignored
Popularity and author do not count. Popularity only decides the order in which skills get tested.
Task bias (limit)
AI writes the tasks from the SKILL.md, so they can lean toward cases where the skill helps. The rubric leaves out rules specific to that SKILL.md.
Few runs (limit)
Two tasks per skill, up to three runs each, stopping early when two runs agree. Treat the score as a guide.
Skills we cannot test
Skills that need an outside account or API key cannot run in isolation, so we skip them.
FAQ
Why do some skills have no score?
They have not been tested yet. We test one a day, most popular first. Every skill still shows the document check from its SKILL.md.
Why does a well-known skill score low?
If Claude Code already handles the task well without it, there is little difference to measure. The score measures what the skill changes, so it is low for work Claude already does well.
What if a score looks wrong?
Use the document check section on the skill page to request a review. The outputs page shows every final message and file from each run.
Which models and dates were used?
Below the hands-on test on each skill page we list the run model, the grading model, and the date.
Recently tested
Skills that went through a hands-on test most recently.
Choose skills by how they actually ran
Sorting by run score puts skills with a proven effect first.