# Our skills, October 2026: 115 measured with and without

> 145 skills joined our catalog in October 2026. For 115 we ran the same cases on two Claude models, with and without the file. Most helped; some did not.

Published 2026-10-28 · https://aiskills402.com/blog/our-skills-october-2026

Most of them helped: 78 of the 115 skills we could compare raised Sonnet's pass rate by more than five points, and 88 did the same for Haiku. For the other 37 and 27 the file changed nothing we could measure, and no skill made either model worse once its checks were pooled. These are the skills we added to the catalog in October 2026, through the 28th, with each one's own test run set against the same request without the file. Our skills are the only ones on this page, and since we sell them, the weak results are listed next to the strong ones.

![A bar chart: for 115 compared skills, 78 helped Sonnet and 88 helped Haiku, none made either worse](https://aiskills402.com/blog/images/our-skills-october-2026/how-many-helped.webp)

## Method

Every skill has a test summary beside its file, with a set of cases we wrote around the mistakes that skill exists to prevent: a repaired document that quietly drops a key, a translated string that loses a placeholder, an image prompt that names a sign with letters on it. Each case ran one time per side on two Claude models, Sonnet 5.5 and Haiku 5.5. Half the runs had the skill's `SKILL.md` loaded; the other half got the identical request alone. The checks are mechanical where a script can decide, and read by hand where it cannot.

To get one number per skill we pooled all of its with-and-without rows into a single pass rate per model. Across the 115 skills that gives 2,685 checks per side. Pooling hides how large each row is, so the chart below shows rates, not row counts. A skill counts as helping when its pooled rate rises by more than five points. The script that does this sits in our repository next to the numbers it produces.

## Results

Pooled over all 115 skills, Sonnet passed 2,650 of the 2,685 checks with the file and 2,251 without it, which is 99 % against 84 %. Haiku passed 2,552 with the file and 2,064 without, 95 % against 77 %. The weaker model gains more because it has more room to lose.

The kind of work decides the size of the gap. The chart groups the skills by catalog category.

![Grouped bars: share of checks passed with and without the skill, for each catalog category, on Sonnet and on Haiku](https://aiskills402.com/blog/images/our-skills-october-2026/by-category.webp)

Marketing text shows the biggest swing on Sonnet, from 42 % to 96 % across three skills, and the single writing skill went from 43 % to 78 %. The 34 code skills moved Sonnet from 89 % to 99 %, so there the file removes the last few misses, not a large gap. Education was flat on Sonnet at 96 % both ways.

## The ten biggest gains

Here are the ten skills where Sonnet gained the most, with the Haiku result as a grey dot.

![Dumbbell chart: Sonnet's pass rate without and with the skill for the ten biggest gains, with Haiku's rate with the skill](https://aiskills402.com/blog/images/our-skills-october-2026/biggest-gains.webp)

The house-style prompt writer ([see its page](/skills/image-prompts-one-style)) is the largest: without it Sonnet met the style requirements in 1 of 24 checks, with it in all 24, and Haiku reached 18. The YAML fixer ([its page](/skills/yaml-repair)) took Sonnet from 16 of 44 to 44 of 44. The text-to-JSON converter and the JSON fixer show the same shape of gap, and for the converter the weaker model reached only 9 of 18 even with the file.

The next group is about text that has to stay exact. The [product copy writer](/skills/product-description) went from 20 of 48 to 44 of 48 on Sonnet, the [app string translator](/skills/translate-app-text) from 9 of 18 to 18 of 18, and the [Bulgarian transliterator](/skills/bulgarian-transliteration) from 15 of 24 to all 24. The list closes with a [changelog writer](/skills/changelog-from-commits) and two review skills, one for [x402 listings](/skills/x402-discovery-listing-review) and one for [Cloudflare claims](/skills/cloudflare-evidence-check).

## Where it changed nothing

For 21 of the 115, Sonnet was already perfect without the file, among them the regex, redaction, RSS feed and SQL conversion skills. Those skills still help the weaker model in many cases, which is why they stay in the catalog, but a reader with Sonnet should expect nothing from them on the cases we wrote. We would rather say that here than let a page of green bars suggest otherwise.

The one row that went the wrong way was in `mojibake-repair`: Haiku restored 14 of 15 texts exactly with the file and all 15 without it. The same skill earned its place in the other row, where six texts had more than one possible original and the skill makes the model refuse to guess. Without it, both models guessed on five of the six.

## What this edition does not include

Of the 145 skills, 30 have no with-and-without comparison. Five are the public-domain book catalogs, which are data files tested by asking questions of the data. Two more, the [done-check skill](/skills/done-means-done) and a prompt-injection guard, carry their own test notes. The remaining 23 are mostly short review skills, such as the Stripe webhook, JWT and Terraform plan reviews, which got a light test of a few cases, one run per model, with the skill only. Each page says what was and was not run.

The September edition used an older Haiku and eight skills; [last month's report](/blog/our-skills-september-2026) is a separate sample and the two should not be added together.

## What we did not measure

- **The pooled pass rates rest on first runs.** A second run could flip a result, so a gap of one check means little. A jump from one check in twenty-four to every check is another matter.
- **Cases we wrote ourselves.** They target known mistakes, which suits a skill built for those very mistakes. Another author would probably find other gaps.
- **Pooled rates.** They weigh a long row the same as a short one, so a skill with one big row can look better than its average case.
- **Mechanical checks.** They count exact words, keys and limits. They cannot see an answer that passes every check and is still not what the reader needed.
- **Two Claude models only.** No other vendor, and no second run on the newer Haiku alias if it moves again.
- **October is not over.** The count stops on 28 October; skills listed after that go into next month's page.