[alekswebnet]

Closes #67.

An agent asked for an H5P quiz could check a package with h5p-verify, but had nothing to build one with. h5p-author is a new skill in skills/, beside the other three. From a brief it:

1. writes a small quiz source;
2. builds a self-contained .h5p from the library pack installed in the user's project;
3. plays it through the verifier;
4. checks the screenshot.

It covers text-only Multiple Choice and Question Set.

Changes

The skill (skills/h5p-author/)

• SKILL.md walks through the brief, the source, the build, the verification and the hand-over, and states the limits:

  • no media and no other content types;
  • the libraries' button labels stay English;
  • a verifier pass covers startup only.

  It tells the agent not to substitute a quiz for another kind of activity, and never to make a check pass by removing a question.

• references/quiz-format.md describes the source format; assets/quiz.json is a two-question starter.

• assets/phishing-scenarios.json, with references/scenario-quiz.md, is a fuller example: eight scenarios, four with several correct answers, an 80% pass mark, an answer key, and attempts that fail at 75%, pass at 83.3% and retry to 100%.

• LICENSE is the repository's MIT licence, installed with the skill.

The helper (skills/h5p-author/scripts/build-quiz.mjs)

• Validation: it checks the source field by field; an unknown field or a wrong type fails with its path.

• Dependencies: it loads @missing-elements/h5p-libraries and @zip.js/zip.js from the calling project, not from the skill's folder, so an installed or symlinked skill works. It refuses a zip.js older than 2.24 and prints the install command. appendZip has no filter option before 2.20, and without it each part's own h5p.json would be copied too.

• Libraries: it takes Multiple Choice and Question Set from the pack and walks their preloaded and dynamic dependencies. Each part is checked against the pack index's SHA-256, and its library.json against the index entry. Only those folders are copied:

  Package	Libraries	Size
  Question Set	10	1.18 MB
  Single question	8	0.94 MB
  Whole pack, for comparison	105	14 MB

• Content: every field comes from the installed semantics.json defaults, and a required field it cannot fill fails the build. Authored text is escaped into the HTML fields. Each question gets a UUID subContentId. h5p.json is written first and content/content.json last.

• Output: it builds in a temporary folder next to the output and copies with COPYFILE_EXCL. An existing file is never overwritten, and a failed build leaves nothing behind.

Tests (apps/skills/, a new workspace app)

The demo is not tested, except its installable app, so the skill's tests live in a project of their own. apps/skills is the kind of project the skill runs in: it installs the pack and zip.js, which the helper resolves from it, and the verifier. Nothing in it ships.

• tests/author.test.ts runs the helper in Node, 29 tests:
  • 21 malformed sources, each refused with its path;
  • both kinds built with the pack's versions, defaults and full dependency set;
  • escaping;
  • no overwrite and no partial output;
  • a damaged part refused by its SHA-256;
  • an older zip.js refused with the install command;
  • the helper run from a copy of the skill outside the checkout, loading its dependencies from the calling project.
• tests/author-verify.test.ts runs both kinds and the scenarios through the verifier's published API, trusted and untrusted against the installed pack (6 tests). Each passes with no warnings and no errors.
• tests/author-playback.test.ts plays the eight scenarios in the verifier's preview, the page the skill will hand a person for review, trusted and untrusted (2 tests):
  • every chosen answer's feedback appears;
  • 9 of 12 (75%) fails the 80% mark, 10 of 12 (83.3%) passes, and a retry reaches 12 of 12;
  • Show solution works;
  • finished comes from the Question Set alone, never from a question.

Docs: the README's agent-skills section, docs/verify.md, docs/development.md, CONTRIBUTING.md, and AGENTS.md's layout and test lists.

Testing

• pnpm typecheck across the workspace passes, the new app included.
• apps/skills: 37 of 37 in 18 s.
• The verifier's own tests, main merged in and the author test moved out: 22 of 22.
• The command line against a stand-in zip.js 2.15: it exits with code 1, prints the install command and "older than 2.24", and writes nothing.
• The playback test runs in Chromium only. While it lived in the demo, the engines job also ran it in Firefox and WebKit, 2 of 2 in each.

Not verified

• An agent taking a real brief through the whole skill. The tests build the committed examples.
• Installing with npx skills add from GitHub, which can only fetch the skill once it is on main. The portable test copies the folder instead.
• Uploading to an LMS. The pack's versions may be newer than the destination's, as SKILL.md says.
• Content in other languages: the libraries' button labels stay English, and the helper warns about it.

Later

The review step will name h5p-verify --preview once h5p-verify 1.2.0 is on npm. npx skills add reads the skill from main, so the step cannot land there before the release.

Release

None. Skills are fetched from the repository rather than published, and the tests don't ship. The README paragraph reaches npm with the player's next release, since prepack copies the root README.

🤖 Generated with Claude Code