Skip to content

Releases

author 0.13.0, h5p-mcp 0.2.0

Choose a tag to compare

@alekswebnet alekswebnet released this 19 Nov 08:48

author 0.13.0

  • h5p-author read: a package back into its source (#236). npx @missing-elements/h5p-author@latest read rivers.h5p rivers writes the source a build takes to make the package again, with its pictures, videos and other files beside it, so you can change it and build it again. readPackage() does the same from code.
    • It reads content made for the library pack's versions of its libraries, and refuses any other, naming both versions.
    • What a source can't hold, such as the author's comments, is noted and left out. What a build would refuse, such as styles pasted from Word, is noted so you can fix it before building.
    • It never writes over a file. An archive is untrusted: only the files its content names are written, and only inside the folder.
  • Notes for what builds but is almost surely a mistake (#241). In the content types where it looks for them, it notes four things: no answer marked correct, two answers or statements alike, a text with no gap, and an optional alt text left blank beside a picture. A build and checkContent both report them, and each note says what to write. H5P allows all four, so none of them stops a build.
  • H5P.Row takes what H5P.com's editor writes (#242): column widths with two decimals, such as 62.56, and a column left empty as a spacer.
  • The skill changes an .h5p you give it (#242). It reads the package with the tool, changes the source as asked and builds it again, instead of editing or zipping the package by hand. The new references/change.md is exported with the other skill files. To update a global install, run npx skills update -g.
  • Dependencies: @zip.js/zip.js ^2.25.0. The package's links now point at the renamed repository, missing-elements/h5p-toolkit.

h5p-mcp 0.2.0

  • h5p_read: change a package you already have (#236). It reads an .h5p back into its source, with its pictures and other files, into a new folder beside the package. The model then changes the source as asked and builds it again with h5p_build. It reads content made for the library pack's versions of its libraries, and says what a source can't hold, such as the author's comments.
  • It needs author 0.13.0, which its dependency range now requires.

Full Changelog: mcp-v0.1.0...author-v0.13.0

h5p-mcp 0.1.0, author 0.12.0, verify 1.6.0, normalize 1.4.0

Choose a tag to compare

@alekswebnet alekswebnet released this 18 Nov 11:48

New: H5P Author MCP, @missing-elements/h5p-mcp 0.1.0

A local MCP server for apps without a shell, such as Claude Desktop. It builds an H5P package of any content type onto your disk, checks that it plays in a real player, and opens it in your browser. It has three tools: h5p_types, which lists the content types and explains how to write a source for one; h5p_build; and h5p_preview.

  • Claude Code: claude mcp add h5p -- npx -y @missing-elements/h5p-mcp
  • Claude Desktop and VS Code: see the README.

Where it reads and writes: only in the folders you give it with --allow. Without those, it uses the folders your app shares with it, and failing that, Desktop, Documents and Downloads.

What it never does: overwrite a file, download from an address on your local network, or send anything to a server of ours.

It needs Node 22 or later, plus Chrome or Edge for the check. It's built on v2 of the MCP TypeScript SDK.

author 0.12.0

  • A source's own fields as a schema: sourceSchema (#230). The build checks type, title, language, license, authors and a source's other fields against it.
    • It's exported, so a tool can show a model the shape of a source before one is written. toJSONSchema from zod 4.2 or later turns it into JSON Schema.
    • Every problem with an author is now reported at once. An author written as a bare name is told the object to write.
    • The Source type is the schema's output, with params optional.
    • New dependency: zod.
  • Videos and recordings are checked by their bytes (#218).
    • A web page saved as clip.mp4 is refused by name, and so is an MP4 without an index a player can read.
    • An MP4 with no sound, with no picture, or encoded in HEVC gets a note.
    • In an Interactive Video, an interaction, bookmark, end screen or start time that falls after the video ends is refused.
  • inside(path, folders): the folder check within applies, for a tool that reads files of its own. It tells whether a path, with links followed, is in one of the folders.

verify 1.6.0

  • A video or recording the browser couldn't play is now a warning (#218): video does not play: <file>: …, with the browser's reason. In a browser without H.264 (or without AAC, for a recording), such as Playwright's Chromium, an MP4 is reported as not checked instead.
  • previewPackage() and openBrowser() are exported, so code can serve a preview the way --preview does.

normalize 1.4.0

  • mp4Info(read, size) from @missing-elements/h5p-normalize/mp4 (#218). It reads an MP4's length and each track's kind and codec from its index, without reading the media. It runs on a file in Node or on a Blob in a page.
  • The faststart rewrite now uses the same reader, which stops after 10,000 atoms, so a file of millions of tiny atoms can't exhaust memory.

Full Changelog: author-v0.11.0...mcp-v0.1.0

author 0.11.0

Choose a tag to compare

@alekswebnet alekswebnet released this 18 Nov 11:48

Confine a build to folders: within

buildPackage(source, output, { within: [folder] }) reads local files, and writes the package, only inside the folders given, judged on real paths, so .. and links out are refused. A path outside gets the same answer whether or not a file is there, a file that cannot be read is a problem of the source's, and the problems name no absolute path. For a tool that builds what a model wrote, such as a local MCP server.

Build without writing: check and build --check

buildPackage(source, undefined, { check: true }) and h5p-author build --check source.json run every check a build runs and write no package. A media file named by web address is still downloaded, to a temporary folder removed after.

Fetch from the public internet only: publicFetch

Pass fetch: publicFetch and a build refuses an address on a local or private network however it is written, IPv4 written as IPv6 included, on every redirect, and drops Authorization, Proxy-Authorization and Cookie on a redirect to another site.

The content as text: outline and outlineOf

Every build result has outline: the content as plain lines for a review, made when it is first read. What the author wrote, the right answers marked, score ranges, slides, chapters and questions numbered, a scenario's links, a video's times; interface texts and settings left out. outlineOf(libraries, content) makes one from what loadLibraries loaded, taking its arguments as checkContent does.

Fixed

  • Markup with a > inside a quoted attribute built into text a browser hides, and the link or tag after it swallowed what followed. It is now refused: write &gt; for it (#222).
  • Tags are read as H5P's content check splits markup, so the check for an empty field and the layouts' measures read a < inside an attribute, or one that starts no tag, as the check does.

Full Changelog: author-v0.10.3...author-v0.11.0

author 0.10.3

Choose a tag to compare

@alekswebnet alekswebnet released this 16 Nov 11:24

h5p-author 0.10.3 changes two things in the skill's text (#211, #213, #215). A chat whose code runs Node now tries to download the build tool whatever its code tool says about the internet, so Grok builds the .h5p in the chat again. And no link carries the source any more: a package built in the chat comes with its source file and the build page's plain address. The build itself is unchanged.

Grok builds in the chat again

Since 0.10.0 the skill took "no internet" from the chat's own tools and then skipped the route that downloads the build tool and the type's libraries from the organisation's site. Grok's code tool is documented as having no network, yet it reaches that site. So Grok on the web handed over the JSON source unless the 15 MB build archive was attached.

Now only a chat with no code tool skips the build. Where Node runs, the skill tries the two downloads even where the tools say there is no internet or npm gave up. If a file does not arrive, it goes on to the build page as before.

  • In Grok, on a free account, with the skill's zip attached and no archive: it built a six-slide Course Presentation in 3 to 4 minutes.
  • In ChatGPT, on a free account, whose code can download nothing: the source arrived in 1 to 2 minutes, and with the archive attached the file built in about 10 seconds.
  • On Haiku, each run told its code tool has no internet: with npm hanging and the web reachable, 4 of 4 downloaded both files and built packages that pass the verifier. With nothing reachable, 3 of 3 tried the download and gave the JSON source, which builds and passes.

No link carries the source

A package built in the chat came with a link to the build page that held the whole source. A command printed the link, and the model copied it into its reply. One wrong character leaves the page nothing to build: Grok handed over a link the build page could not read, as ChatGPT had done before.

So the link is gone. The reply now holds:

  • the .h5p file;
  • its source, as <name>.source.json;
  • one fixed line: To play it, change it or build it again: open https://missing-elements.github.io/h5p/build/, choose <name>.source.json under "Or open a file", and press Build.

Where the source names files, the reply adds that the page asks for them. A chat that built nothing already handed over the source this way since 0.10.2. The build page still opens #source= links, which copies of the skill already installed print.

On Haiku, in three places (npm working with the verifier, the two downloads, the archive attached), 3 of 3 replies gave the line whole with the file's name and no link, and every package passes the verifier. Not yet measured: Grok and ChatGPT with this hand-over.

Getting it

A copy installed with npx skills add takes the change with npx skills update -g. A copy uploaded to Claude or ChatGPT 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. Grok has no place to install a skill: attach the zip to the message where you ask.

Full Changelog: author-v0.10.2...author-v0.10.3

author 0.10.2

Choose a tag to compare

@alekswebnet alekswebnet released this 15 Nov 01:10

h5p-author 0.10.2 changes how a chat that built nothing hands the source over (#208, #209). It now gives the source as a JSON code block, or as a file where it can make one, with the build page's plain address. No link carries the source. The build itself is unchanged.

Why the link went

0.10.1 had the model write the build page's link itself, with the source in it as readable JSON. Haiku wrote working links, but ChatGPT on a free account did not:

  • Its model escaped characters it did not need to, then wrote spaces raw partway through: "question":"What happens during evaporation?". A Markdown link cannot hold a raw space, so ChatGPT showed the address as text, and the link it made stopped at the first space, after 283 characters of the source.
  • The address in angle brackets, […](<…>), which lets a Markdown link hold spaces, came out of ChatGPT as [Build and play it](: its renderer drops the <…>.

A link that carries the source needs every character of a long text right, and a chat app's model does not always manage that.

What the reply holds now

  • The whole source, once, in one json code block, which the chat's copy button copies exactly.
  • One line: copy it, open https://missing-elements.github.io/h5p/build/, choose Paste your chat's JSON, paste it in place of the sample's, and press Build. Where the source names files, the page asks for them.
  • Only where the chat has a code tool: the source as a .source.json file too, which the page opens under Or open a file, and the line offering the build archive.

Outside a chat, the file takes the code block's place. Routes 1 to 3, which build the package in the chat, keep the compressed link their command prints. The build page still opens #source= links of both kinds.

Measured

Haiku, the weakest model first, reading the skill through a logging reader. 6 of 6 replies held one json block that built and passed the verifier, the plain address, and no link carrying the source. They covered Question Sets in English and Ukrainian, an Accordion about HTML tags, a True/False naming a photo, and a Drag the Words with % and parentheses. Both runs with a code tool and no network also gave the .source.json file, equal to the block, and the archive line, and none of the four without one did.

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-v0.10.1...author-v0.10.2

author 0.10.1

Choose a tag to compare

@alekswebnet alekswebnet released this 15 Nov 00:26

h5p-author 0.10.1 changes how a chat that built nothing hands the source over (#206, #207). The model now writes the build page's link itself, with the source in it as readable JSON. The reply has no code block, and the person clicks one link. The build itself is unchanged.

One link, no code block

In 0.10.0 a chat that could not build wrote a small HTML page in a code block. The person opened it with the chat's Preview, waited, and clicked the link the page made. That worked in ChatGPT on a free account, but it showed the person a block of code and took three steps.

A model cannot compress a source, which is why the page's script did it. It can write its own JSON, though. Since missing-elements/missing-elements.github.io#54 the build page also reads the source as JSON written straight into the link, so the skill's route 4 is now one Markdown link:

[Build and play it](https://missing-elements.github.io/h5p/build/#source={"type":"H5P.TrueFalse",…})
  • The source is on one line, with eight characters escaped so the link holds: a space, %, &, \, (, ), < and >. Everything else is written as it is, letters of any language included.
  • A source that names files gets the link too: the build page asks for the files before it builds.
  • Where a code tool runs, the reply also offers the build archive, as before.
  • Outside a chat, where a terminal may not open such a link, the source is also saved as <name>.source.json, which the build page opens.
  • Routes 1 to 3, which build in the chat, keep the compressed link their command prints.

The build page reads the link leniently, because a weak model's JSON slips. It closes up to three braces or brackets left open at the end, and drops what follows the last } where a link runs on into its sentence. JSON it still cannot read goes in its box with where it breaks, and a button copies that for the chat to fix.

Measured

Haiku, the weakest model first, read the skill through a logging reader. Each reply was rendered as a chat renders Markdown, its link decoded as the build page decodes it, then built and verified:

Built and passed
First wording 5 of 7
Released wording 12 of 12

The 12 were Question Sets in English and Ukrainian, Accordions about HTML tags, True/False questions naming a photo, and a Drag the Words with % and parentheses, some from a code tool with no network. No reply held a code block. Links ran from 0.5 to 6.9 KB as written. On the live site, the links Haiku wrote built and played, or asked for the photo, in Chromium. The page's eleven kinds of test link did what they should in Chromium, Firefox and WebKit.

Not yet measured: how ChatGPT and other chat apps render a link of several KB.

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-v0.10.0...author-v0.10.1

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

author 0.9.8

Choose a tag to compare

@alekswebnet alekswebnet released this 13 Nov 20:50

h5p-author 0.9.8 lets a chat build from an archive the person attaches, where its code can download nothing (#200, #201). It also drops the skill's two passages for a chat that has its text alone (#196, #197), and fixes --version in the command line as one file (#198, #199). Every source builds as before.

Build from an archive the person attaches

ChatGPT's sandbox runs Node, but nothing from outside reaches it. npm goes through OpenAI's mirror, which fails or hangs. Its download tool saved none of eleven files tried, from our site or any other, and an installed plugin's files never reach its code. A file the person attaches does: with the build tool and the libraries attached as a .tar.gz, ChatGPT built a Question Set that passes the verifier. Claude and Grok built from the same archive.

The archive is h5p-author-build.tar.gz, 14.9 MB: the one-file build tool, every library in the pack, and the libraries' licences. The organisation's site writes it from the release it pins, at https://missing-elements.github.io/h5p/bundles/h5p-author-build.tar.gz, so it holds this release from the site's 0.9.8 pin on. It is a .tar.gz because ChatGPT took the same files as a .zip for a plugin to install and refused it.

  • Step 3 builds from an attached archive first: unpack it, then node h5p-author.mjs build source.json course.h5p --libraries parts. This needs nothing installed or fetched, for any type. With no network, a picture or other file must be one in the sandbox: a web address is downloaded as the package builds.
  • Where Node runs but nothing installs or downloads, the build page's hand-over carries one more line, right after the link: To get the .h5p file here instead: download https://missing-elements.github.io/h5p/bundles/h5p-author-build.tar.gz (15 MB) and attach it to your next message.

Nothing passes through a server of ours on this route. The archive is the public tool and libraries, and the chat's provider sees no more than the chat already gave it.

The skill is installed

The skill no longer tells a chat that has its text alone to open its page at /h5p/skill/ and take the addresses from there. Since 0.9.7, every chat it is offered to takes it installed: from a zip in Claude, ChatGPT and Grok, or with npx skills add. Step 3 still takes the build tool's and the bundles' addresses from that page, which stays.

The command line as one file prints its versions

h5p-author --version in the one-file tool the site serves crashed: it read package.json files that are not beside that file. It now reads them where the package is installed, and in the one file uses copies the bundler writes in (about 9 kB). A model that checks the tool with --version first no longer sees it fail. ChatGPT's plugin creator found the crash.

Measured

Haiku, the weakest model first, each run in a sandbox with npm cut off:

Old text New text
Archive attached: built from it, package passes the verifier 3 of 3 3 of 3
No archive: the line offering it, exact — 2 of 3
No archive: the link exact, the source file given — 3 of 3

Haiku already builds from an attached archive under the old text, so the new route is for models that need the step written out, as the owner's ChatGPT prompt wrote it out. One run in three left the offer out. So the guide tells people using ChatGPT to attach the archive to the message where they ask.

Without the text-alone passages, 6 of 6 Haiku runs still reached the build tool and the bundle through the page and built.

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-v0.9.7...author-v0.9.8

author 0.9.7

Choose a tag to compare

@alekswebnet alekswebnet released this 13 Nov 13:30

h5p-author 0.9.7 changes the skill's text only (#194, #195). Where a chat cannot build the package, it now hands over the source as a file beside the build page's link. The skill also says that only the build tool writes an .h5p, and that an npx which hangs has failed. The build does not change: every source builds as before, and no problem reads differently.

The source as a file, beside the link

Where nothing is built in the chat, step 3 ended with the build page's link alone. A chat can get characters wrong when it writes a long link into its reply, and a compressed source does not survive one wrong character: ChatGPT's link was unreadable in 2 of 3 tries, while the source file it gave built every time.

So where a chat can give the user a file, it now gives the source too, named after the activity with .source.json (course.source.json). The build page opens it under "Or open a file". Since 2026-11-13 the page also says when a link arrived damaged, with a button that opens that file.

Only the build tool writes the .h5p

Pressed for "the file itself" where it could not install the tool, ChatGPT zipped h5p.json and content.json by hand: a 2 kB archive with no libraries in it, which fails the verifier. The skill's opening now says never to put an .h5p together yourself.

An install that hangs has failed

Step 3 listed the errors that mean npm cannot be reached: EAI_AGAIN, ENOTFOUND, a registry's 403, 404 or 503. In ChatGPT and in Grok, npx neither failed nor finished: it timed out. The list now names an install that hangs until the command times out, so the skill moves on to the two downloaded files or to the build page.

Measured

Haiku, the weakest model first, 22 runs:

0.9.6 0.9.7
No npm: the exact link and the source file 3 of 3 3 of 3
Asked for "the .h5p file itself", not a link: the exact link and the file 1 of 3 2 of 3
The same request: ended with an npx command for the user 1 of 3 1 of 3
A registry that never answers: moved on to the two files and built 2 of 2 2 of 2
A hand-made .h5p 0 of 6 0 of 6

Haiku already did what the new sentences ask, so the case for them rests on ChatGPT's replies. One request is not covered: asked for a file and not a link, 1 run of 3 under each text ended with an npx command for the user to run.

The 0.9.6 runs were only told that npm could not be reached. Three 0.9.7 runs that were only told so used the machine's npm and built, so 0.9.7's six were run again with npm cut off.

Grok, with 0.9.6 installed from its zip, built a Question Set from the two downloaded files after its npx timed out (2026-11-13). It is the first app seen taking that route.

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.

Full Changelog: author-v0.9.6...author-v0.9.7

author 0.9.6

Choose a tag to compare

@alekswebnet alekswebnet released this 13 Nov 08:34

h5p-author 0.9.6 changes the skill's text only (#192, #193). A content type with no type file can now be read where npm cannot install, and a finished reply carries one web address, the build page's link. The build does not change: every source builds as before, and no problem reads differently.

A type with no type file, where npm cannot install

27 of the 55 content types the site bundles have no type file, among them Crossword, Find the Words, Drag and Drop and Flashcards. For those, step 1 said to run npx @missing-elements/h5p-author@latest schema <Type>, which fails where npm cannot install. It now says to fetch the build tool and the type's bundle, as step 3 does, and run:

node h5p-author.mjs schema <Type> --libraries parts

One web address in the hand-over

Step 3's two commands print the build page's link alone, and for a source that names files they print why there is none:

No link: the source names files, which a link cannot carry.

Step 5 hands over the package, its source file and that link, and no other web address. It no longer hands over 0.9.5's line on where to say how it went, the libraries.txt link, the tools to go to next, or report.json and its revision. The libraries' licences stay as a sentence and the screenshot stays; the preview command goes over only where there is no link. Where nothing is built in the chat, the build page's route is the whole hand-over: the link, or the whole source and the page's address.

The build page still asks how it went at its foot, under the id feedback-heading that installed copies of 0.9.5 link to.

Measured

Haiku, the weakest model first. Crossword and Find the Words, read by the skill's link with no package registry, two runs each:

0.9.5 0.9.6
Fields at the first schema call 0 of 4 4 of 4
Built a package 3 of 4 4 of 4, all at the first try

All 7 packages pass h5p-verify. With the skill, a shell and npm, the same requests as 0.9.5's measurement:

0.9.5 0.9.6
Three-question Question Set: web addresses in the reply 2 or 3 1, the link, decoding to the source exactly
Multiple Choice with the user's picture: web addresses 1 or 2 none, and the preview command
Question Set reply length 3,146 and 3,724 characters 2,250 to 2,737 characters

As a chat with no tools, 5 of 5 replies gave the whole source and the page's address, and no preview command. None of the replies with a shell said where to tell how it went: that now rests on the page's own section.

A copy installed with npx skills add takes the change with npx skills update -g; a copy uploaded to the Claude app needs a new upload. https://missing-elements.github.io/h5p/skill/ serves this release.

Full Changelog: author-v0.9.5...author-v0.9.6