RSS Feed Builder: test results

Verdicts

Date

2026-10-08

Strong · claude-sonnet-5-5 (Claude Code alias "sonnet")

Right on all 22 cases, scored by program: dates in the mail format with the right weekday, including a time with a US offset that falls on the next day in UTC; no date for the post that had none and no build date when nothing was dated; no description for a post without a summary; no feed link invented when no address was given; ampersands escaped once, HTML summaries in CDATA; the permanent id used as guid and marked as not an address; a percent-encoded address kept as written; the Pinterest items got an image tag without a made-up length; an instruction planted in a summary stayed as text.

Weak · claude-haiku-5-5 (Claude Code alias "haiku")

Right on 21 of 22, but still missed one: in the case that asked for the three newest posts it wrote the right posts and times and gave wrong weekdays to the dates (a Tuesday written as Sunday), which strict validators reject. The other 21 cases, including Pinterest, ampersands, guids, undated posts and the planted instruction, were correct.

With and without the skill

Tested 2026-10-08.

The same request on both sides, XML only asked for in both; a fence around the answer is removed first. Sonnet without the skill was right on all 22, so for Sonnet this skill buys nothing measurable here. Haiku without the skill got 20: it put the Pinterest pictures into the description as escaped image markup instead of a media tag, and in the planted-instruction case dropped the summary that held the instruction and left out the feed's self link. With the skill Haiku fixed those two but got the weekdays wrong in the newest-three case, which it had right alone. The gain is for the weaker model, and it is small: two cases fixed, one lost.

Same cases and the same checks with and without the skill. The cases are ours, written around what the skill is for; with a handful of cases, a difference of one or two is within noise.

Note

Twenty-two cases written by us, one in Bulgarian and one in Atom: plain feeds, a feed asked to keep only the three newest posts, posts without a date, summary or picture, ampersands in titles and addresses, an HTML summary and an HTML title, permanent ids, a percent-encoded address, an author given only as a name, a missing feed address, a relative-path list, a Pinterest feed with and without byte sizes, and an instruction planted in a summary. There are no XML checks in our tester, so each answer is scored with regular expressions that accept every valid form: either quote style, any attribute order, CDATA or entities, a date in its own offset or in GMT, no weekday at all. Only the invented or wrong fact stays strict: ISO dates, a wrong weekday, an invented date, description, picture or feed address, a raw ampersand or raw HTML, a changing guid, an enclosure with a made-up length. A first version rejected an HTML title written with escaped tags and a namespace declared on the link element itself; both are valid and were relaxed after the run, and the numbers below are from rescoring the recorded answers. The Pinterest rules (four required fields, media tag when the size is unknown) are owner-measured in 2026 and were not re-checked against Pinterest's pages. One run per model and case.

What was not measured

• Models other than the two named above were not run.
• Each verdict comes from the test run on the date shown; the skill may have changed since (check the version).
• Full test inputs are not published here, only short excerpts of our own text.
• Results on your own texts, languages and domains can differ.