Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -136,9 +136,9 @@ npx skills add missing-elements/h5p-offline-player
| [`h5p-player-setup`](https://github.com/missing-elements/h5p-offline-player/tree/main/skills/h5p-player-setup) | Puts H5P on a website, from the player on the site itself or framed from a domain of your own |
| [`h5p-normalize`](https://github.com/missing-elements/h5p-offline-player/tree/main/skills/h5p-normalize) | Rewrites a package whose video will not stream |

A chat that cannot install a skill can be given `h5p-author` as a link instead:
[missing-elements.github.io/h5p/skill/](https://missing-elements.github.io/h5p/skill/) lists each
of its files at its full address, and each file its build fetches where npm cannot be reached,
In a chat app, `h5p-author` installs from a zip: Claude and Grok take one, ChatGPT's Plugins
another, both at
[missing-elements.github.io/h5p/create-with-ai/#install](https://missing-elements.github.io/h5p/create-with-ai/#install),
from the release the build page runs.

## Contributing
Expand Down
50 changes: 36 additions & 14 deletions design/author.md
Original file line number Diff line number Diff line change
Expand Up @@ -306,16 +306,24 @@ to 3,000. Twenty-four things the code does not say by itself:
the letter. An uploaded plugin is no way round the registry: a ChatGPT plugin that carried the
one-file tool and the whole pack, 15.6 MB, uploaded, and its code's `find /` found no file of
it, so the zips the organisation's site offers to install (missing-elements.github.io#40)
carry the skill's text and files only.
carry the skill's text and files only. Nor is ChatGPT's own download tool, on the owner's free
account the same day, asked step by step: its web tool opened the site's text, JSON and
Markdown, `SKILL.md` whole, and refused the build tool, a bundle and a `.h5p` by their content
type (`text/javascript`, `application/x-zip-compressed`, `application/octet-stream`), and its
`container.download` saved none of eleven files, four from other sites among them, Simon
Willison's own test file too, which he had downloaded with it in January 2026. So in ChatGPT
the build page is the route, whatever type the site serves a file as.
- **A bundle per content type, where the pack cannot be installed** (2026-11-09, #166). The page
leaves the last steps to the user, but ChatGPT's sandbox runs Node and fetches a file its
browsing tool has opened: given the libraries as files, a chat builds and fixes its source
itself. The skill chooses the content type first, so the unit is the type. `bundleLibraries` is
the build's own rules over the most a source of the type can name: the type and every library
its library fields offer or give as their default, theirs in turn; `planPack` over all of them
as roots, as `closure` in `build.mjs` plans one package; `installClosure` over what that plans.
Both only ever add for more roots, so one source's libraries are within its type's bundle.
`writeBundle` zips them as `parts/`, with `libraries.txt`, without which the parts are not handed
itself. Grok's sandbox did so on 2026-11-13; ChatGPT's, on the owner's account that day, did not
(*Where `npx` cannot install, a page builds*). The skill chooses the content type first, so the
unit is the type. `bundleLibraries` is the build's own rules over the most a source of the type
can name: the type and every library its library fields offer or give as their default, theirs in
turn; `planPack` over all of them as roots, as `closure` in `build.mjs` plans one package;
`installClosure` over what that plans. Both only ever add for more roots, so one source's
libraries are within its type's bundle. `writeBundle` zips them as `parts/`, with
`libraries.txt`, without which the parts are not handed
on. The bundle carries the pack's whole index, not its own entries: a build from it plans as it
would from the pack, so it writes the same package (the skill's 32 examples were built both ways
and matched, but for the ids a build draws) or stops and names the part it lacks, where a cut-down
Expand Down Expand Up @@ -399,10 +407,10 @@ to 3,000. Twenty-four things the code does not say by itself:
organisation's site serves the skill from the release its build page pins, at
`https://missing-elements.github.io/h5p/skill/`, a page that writes out each file's full address,
since a chat's fetch tool may open only addresses it has seen whole and `SKILL.md`'s relative
links are not (missing-elements.github.io#15). The skill sends a chat to that page for every
address (*A chat with the skill's text alone takes every address from the skill's page*, below),
not to raw.githubusercontent.com. Installing the skill stays the way: nothing is fetched while
the chat works.
links are not (missing-elements.github.io#15). Step 3 takes the build tool's and the bundles'
addresses from that page (*The skill is installed, and its page lists what step 3 fetches*,
below), not from raw.githubusercontent.com. Installing the skill is the way: nothing is fetched
while the chat works.

The same test on the 23 other content types the skill lists says how far each thing reaches:
two runs each, 46 first replies, each built package checked against what its task asked for.
Expand Down Expand Up @@ -471,9 +479,9 @@ to 3,000. Twenty-four things the code does not say by itself:
as written, does not. `sourceType(source, names)` is the build's reading of a source's type,
exported for it (#176), so that the page hands over the type the build took and keeps no copy
of the rule: on the 328 stored sources it names what `sourceOf` does in every one.
- **A chat with the skill's text alone takes every address from the skill's page** (2026-11-10,
#180). An outside review of the skill, by a model given its link, found that `SKILL.md` gave such
a chat addresses to put together: the brief as a relative link, a type's file as
- **The skill is installed, and its page lists what step 3 fetches** (2026-11-10, #180;
2026-11-13). An outside review of the skill, by a model given its link, found that `SKILL.md`
gave such a chat addresses to put together: the brief as a relative link, a type's file as
`…/h5p/build/types/<Type>.md`, a bundle as `…/h5p/bundles/<type>.zip`; its own fetch tool refused
`…/bundles/H5P.Blanks.zip` put together from the last. `/h5p/skill/` wrote out the skill's files
whole for this (*The package carries the skill's type files*, above), but `SKILL.md` never named
Expand All @@ -490,6 +498,20 @@ to 3,000. Twenty-four things the code does not say by itself:
play, and the third reached npm and built with it. The first three had the tool swapped for a
stand-in that logged its runs, and each refused to run it, as it should: a test serves the file
as it is.
On 2026-11-13 the owner dropped reading the skill by its link as a way to give it to a chat, and
the organisation's site stopped offering it (missing-elements.github.io#42): no app tested
needed it. Claude and ChatGPT take the skill as a zip, and Grok takes Claude's, under Skills,
then Import; Gemini opens no address, and a chat whose only web tool is a search gets snippets.
So `SKILL.md` no longer tells a chat with its text alone to open the page first. The page stays,
and step 3 still takes the build tool's and the bundles' addresses from it: installed copies of
the skill name it, and Grok, its `npx` timed out, built from those two files that day, the
first app seen to. Measured on Haiku in the chat stand-in above, the user's message giving the
page's address as the old prompt did, three runs each way: under either text all six opened the
page, the brief and the type files, fetched the tool and the bundle, and built a package the
verifier plays, since the page itself says to read `SKILL.md` first and lists every address.
With the skill installed and npm cut off, both runs handed over the link and the source file,
and one dropped two characters of the link while writing its reply, as ChatGPT had: the first
time on Haiku, and the source file still builds.
- **A source's own fields are written where a chat reads, and the author is asked for**
(2026-11-11, #182). The owner's run in ChatGPT on 0.9.1 wrote a quiz with no `authors`. Each
of the 28 content types the skill lists refuses a source without `title`, `language`,
Expand Down
8 changes: 5 additions & 3 deletions packages/author/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -137,9 +137,11 @@ The package carries those type files too, as the release has them:
these fields. A page that builds for a chat with no skill beside it hands them over with a source's
problems, as the [build page](https://missing-elements.github.io/h5p/build/) does. The skill's text
travels with them, `@missing-elements/h5p-author/SKILL.md` and its brief,
`references/authoring.md`, for a site to serve to a chat given a link to the skill, as
[missing-elements.github.io/h5p/skill/](https://missing-elements.github.io/h5p/skill/) does:
GitHub limits the requests of an address many users share, as a chat app's is.
`references/authoring.md`, so a site can serve the skill from the release it pins:
[missing-elements.github.io](https://missing-elements.github.io/h5p/create-with-ai/#install)
packs them as the zips chat apps install, and lists the build tool and each bundle that the
skill's step 3 fetches at
[missing-elements.github.io/h5p/skill/](https://missing-elements.github.io/h5p/skill/).

## A content type's bundle

Expand Down
9 changes: 1 addition & 8 deletions skills/h5p-author/SKILL.md
Original file line number Diff line number Diff line change
Expand Up @@ -14,10 +14,6 @@ commands for the user to run. Only it writes the `.h5p`: never put one together
`h5p.json` and `content.json`, since one made by hand carries no libraries and nothing checks it.
Keep [LICENSE](./LICENSE) with the skill.

With this text alone and none of its files beside it, open https://missing-elements.github.io/h5p/skill/
first and take every address from it: it writes out in full each file this text names and each
file step 3 fetches, and a fetch tool may refuse an address put together from a name.

Work in this order, one question at a time: the content type, chosen with the user; then the
author and licence, which they give or skip; then the source; then the build, the best way this
environment allows.
Expand Down Expand Up @@ -77,10 +73,7 @@ the same package.
## 2. Write the source

Read the type's file, `references/types/<Type>.md`: notes, an example, its fields. Parts other
content holds have files too: H5P.AdvancedText, H5P.Image, H5P.Video, H5P.Text. With this text
alone and no file beside it, open the type's file at the address
https://missing-elements.github.io/h5p/skill/ lists for it; where nothing opens, write the source
from what you know of the type: the build says, for each place that is wrong, what to write there.
content holds have files too: H5P.AdvancedText, H5P.Image, H5P.Video, H5P.Text.

```json
{
Expand Down
Loading