Skip to content

author 0.10.0

Choose a tag to compare

@alekswebnet alekswebnet released this 14 Nov 22:26

h5p-author 0.10.0 makes the build-page link in a chat that built nothing from a page the chat writes, so no compressed text is typed or copied (#202). It stops a chat waiting on a verifier that npm cannot fetch (#202). And it splits the skill into one file per build route and asks one question at most, the content type (#204). Every source that built still builds the same way: of the build's code, only the text of SOURCE_FIELDS changed.

In a chat that built nothing, a page makes the link

The link to the build page was the weak point of a chat that cannot build. A command printed compressed text and the model copied it into its reply, and ChatGPT broke 2 links of 3. With no code tool, as on a free account for hours, there was no command and no link at all.

The skill's fourth route now has the chat write a small page in an html code block. The page holds the source in a <noscript> and has a fixed script that deflates the source and writes the link, Build and play it. The reply asks the user to open the block with Preview and click the link. Where the chat shows no Preview, the user saves the block as an .html file and opens it. Outside a chat, the skill writes the page to an .html file. A source that names files gets no page: it goes over whole in a json block, and its files are chosen on the build page.

The <noscript> matters for sources about HTML. A <script> holding the source ended at the first </script> in its text, so a Drag the Words that showed the tags made a link to half its source. With scripts on, a <noscript> ends only at </noscript, so nothing in a source has to be escaped by the model.

The verifier is asked for, not waited for

After a build from the archive, ChatGPT ran the verifier's npx, which cannot install there, and waited until its tool gave up. The skill now asks npm for the verifier's version once, with a 10-second limit, early and beside its first file reads, since the check waits on nothing. It runs the verifier only where npm answered. Against a registry that takes a connection and never answers, the check gives up after 10.5 seconds. Where npm cannot reach its registry, the package is handed over unverified, and the reply says so.

One question, one file per route

SKILL.md held all four ways to build, 18.9 KB that every chat read whole to take one of them, and it said some things two or three times. Now:

  • SKILL.md is the flow: the content type, the source, two answers found once (is there internet, does npm reach its registry), then the first of four routes that holds.
  • Each route is a file in references/build/: archive.md (the attached build archive), npm.md, two-files.md (the tool and one bundle fetched), browser.md (the page). The three that build hand over through built-here.md: the build's problems, the file name, the verifier, the link.
  • The author and the licence are no longer asked. What the user gave is used as given. An author not given is the name the model knows them by, or "Unknown". A licence not given, or one no H5P code stands for (such as MIT), is U. The reply says which, so the user can change them. This rule and the licence codes are written once, in SOURCE_FIELDS, under every content type's fields. Only the six CC BY codes take a version, and CC0 1.0 is one code.
  • A file name already taken is not asked about: the skill takes a new name and says so.

A chat now reads 9.8 KB of the skill's own text in the user's browser, 11.3 KB with npm, 11.5 KB from the archive and 12.0 KB from two files, where every route read 18.9 KB before: 37 to 48% less.

For the package

  • A new export, ./references/build/*.md, for the five route files. prepack copies all of references/.
  • SOURCE_FIELDS has the new author and licence text. It is the sentence every type file carries under its fields, and all 28 generated type files carry it.

Measured

Haiku, the weakest model first. Each run read the skill through a reader that logs the files it serves, and the network was cut for real where a run had none:

Result
No tools, or no network: a page written, its link carries the source, the source builds and passes the verifier 24 of 24 pages
The <noscript> page copied as given; both sources about HTML (<script>, </script>, <!--) built and passed 3 of 3
"A quiz": one question asked, the type 1 of 1
Author and licence not given: none asked, "Unknown" and U written and said every run
Licences given, one of them with no H5P code: written as the build takes them, each built 3 of 3
Archive attached and a registry that hangs: built, one connection, handed over unverified 2 of 2
npm working: built, verified, link to the delivered source 2 of 2

In ChatGPT on a free account, Preview showed the page, and its link opened the build page with the right source. That run used the page in its first form. The <noscript> form has not been run in ChatGPT yet.

Where a code tool says its sandbox has no internet, Haiku still ran the 10-second npm check in 4 of 6 runs, and reached the same route. Stronger models were not measured on this.

A copy installed with npx skills add takes the change with npx skills update -g. A copy uploaded to Claude, ChatGPT or Grok needs a new upload, from the zips at https://missing-elements.github.io/h5p/create-with-ai/#install, which hold this release once the site pins it.

Full Changelog: author-v0.9.8...author-v0.10.0