agentsclimarketplace

Generate explainer html

Skill hirokita117/yaml-to-html-skill/skills/generate-explainer-html

turn core.yaml + view.yaml (the explainer's meaning + presentation strategy) — plus quiz.yaml when present — into a self-contained html bundle that helps users understand documents, repositories, pull requests, design notes, or specifications. the bundle has a light/dark shell whose right pane switches between multiple iframe view documents, and copyable prompt templates that ask a local-file-reading ai to add another switchable view. an interactive quiz tab (comprehension check with instant grading) is added by default unless the user explicitly declines. use when the user has (or wants) a core.yaml/view.yaml pair and wants a visual html explainer, or wants to add a new representation (table, worktree, cards, faq, beginner, engineer, pdm view, quiz, …) to an existing bundle. produce the yaml first with the generate-explainer-yaml skill.From its SKILL.md

Install
npx -y skills add hirokita117/yaml-to-html-skill --skill generate-explainer-html

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

10.3 KB, as published. Nobody here has run it

generate-explainer-html

Turn a core.yaml (meaning) + view.yaml (presentation strategy) pair into a self-contained HTML bundle that:

  • shows a tailored explanation UI inside an iframe, with a view switcher in the right pane so the reader can flip between multiple representations, and
  • offers copyable prompt templates outside the iframe so the user can ask another local-file-reading AI to add another switchable view.

This is the second half of the pipeline. The core.yaml / view.yaml come from the generate-explainer-yaml skill (or the user). This skill reads them by absolute path, authors iframe view documents, and assembles the bundle.

core.yaml + view.yaml (+ quiz.yaml)  (from generate-explainer-yaml)
  ↓ design + generate (this skill)
HTML bundle: index.html (light/dark shell + view switcher + prompt cards)
             + views/NN-<id>.html  (one switchable iframe view each, incl. a quiz tab)

The bundle (the output)

The output is a directory, not a single file:

<bundle>/
  index.html          shell: header (theme toggle) + left prompt pane + right view switcher + ONE <iframe>
  views.json          ordered manifest of views {id, label, file}
  core.yaml           copied in (its absolute path is cited by the prompts)
  view.yaml           copied in
  quiz.yaml           copied in when present (comprehension-check questions)
  views/
    01-<id>.html       an iframe view document (full <!DOCTYPE html>, light default)
    02-<id>.html

The right-pane iframe loads one view at a time via src="views/<file>"; the tab switcher swaps it. Views are additive — each new one becomes another tab; existing views and the shell/prompts are never rewritten.

Do not hand-edit index.html or the prompts — scripts/build_html.py is their only author. You author iframe view documents only. Do not build a web app, server, API, or chat UI, and do not depend on any external CDN/CSS/JS.

Steps

  1. Get the YAML files. Take the absolute paths to core.yaml, view.yaml, and (by default) quiz.yaml (from the generate-explainer-yaml skill, or the user). Read them. If they don't exist yet, produce them with generate-explainer-yaml first.

  2. Author one iframe view document — a full <!DOCTYPE html> document tailored to the target and reader. Choose the form(s) from view.yaml (table, worktree, cards, faq, comparison, sequence, reading path, risk/dependency/impact map, glossary, tutorial, review checklist, …). It must include source references, important concepts, relations, what to read next, and next questions, and have real visual structure. It must be light by default and read #theme from its own URL hash (load + hashchange). Inline CSS/JS only; no network. Follow references/html-generation-rules.md. Example: references/sample-iframe.html.

  3. Author the quiz view document (default — skip only on explicit opt-out). Unless the user explicitly declined a quiz, also author an interactive quiz view from quiz.yaml (or, if no quiz.yaml exists, derive items from core.yaml): answer → instant grading → explanation, order adapted to the view.yaml audience and per-item difficulty. Rules: references/quiz-view-rules.md. Example: references/sample-quiz-iframe.html.

  4. Author the prompt templates as a prompts.json array — the "add a view" cards (table, worktree, beginner, engineer, PdM/Biz, quiz, free-form). Each cites the YAML via {{core_yaml_path}} / {{view_yaml_path}} (the quiz card also {{quiz_yaml_path}}) and must not embed YAML content. Patterns: references/prompt-template-patterns.md. Starter set: references/sample-prompts.json.

  5. Build the bundle with scripts/build_html.py (see below). Pass --core / --view / --quiz (copied in; their absolute paths feed the prompt placeholders) and one or more --view-html "ラベル=その.html" — including the quiz view, e.g. --view-html "クイズ=quiz.html".

  6. Validate the bundle with scripts/validate_html.py on index.html and every views/*.html. Fix any error and re-validate until it exits 0.

  7. Hand the user the bundle folder with open instructions (below). Summarize which views exist and which "add a view" templates are available.

Adding a view later (the additive flow)

To add a representation next to the existing ones — do not overwrite:

  1. Author one more iframe view document (or have a local-file-reading AI do it via a copied prompt — those prompts already say "read the YAML at the given path; return one new view").
  2. Re-run the build with only the new view:
    python scripts/build_html.py --bundle <dir> --prompts <prompts.json> \
      --view-html "テーブル=table.html"
    
    It appends the view to views.json and regenerates index.html with a new tab. The existing views, the shell, and the prompts are preserved. (Re-running with an existing label updates that view in place instead of duplicating it.)

Using the scripts

Note: invoke with python3 if python is not on PATH.

Build / grow the bundle:

python scripts/build_html.py \
  --bundle ./explainer-bundle \
  --core /abs/path/core.yaml \
  --view /abs/path/view.yaml \
  --quiz /abs/path/quiz.yaml \
  --prompts prompts.json \
  --view-html "エンジニア=engineer.html" \
  --view-html "クイズ=quiz.html"

--bundle and --prompts are required; --core / --view / --quiz are optional (recommended — the prompts cite their absolute path) and are left in place on a re-run if omitted; --view-html is repeatable and appends/updates views.

Validate (safety / self-containment check), including every view file:

python scripts/validate_html.py ./explainer-bundle/index.html ./explainer-bundle/views/*.html

build_html.py is an assembler, not a renderer: it never turns core.yaml into a fixed diagram. The understanding UI is the iframe view document(s) you authored in step 2.

Opening the bundle

  • Chrome/Edge block file:// iframe loads. Open index.html in Firefox over file://, or serve the folder with a trivial local static server (e.g. python3 -m http.server run inside the bundle dir — it reads only local files; the no-network rule constrains the content, not how you serve the folder).
  • The right-pane tabs switch views; the header toggle flips light/dark for the shell and propagates the theme into the iframe via the #theme hash.

Troubleshooting

  • validate_html.py fails on an http:// / https:// string — a URL leaked into a source_ref or a prompt body. URLs are labels here, not live links: drop the scheme (write example.com/path) or remove the URL, then re-validate.
  • validate_html.py fails on fetch( / XMLHttpRequest / localStorage / etc. — the forbidden token appears literally inside a prompt body or a view. Reword it generically; see references/prompt-template-patterns.md → "Safety wording inside prompt bodies".
  • Iframe is blank in Chrome over file:// — expected; use Firefox or serve the folder.
  • build_html.py: --view-html must be "LABEL=PATH" — pass each view as ラベル=パス.
  • build_html.py: no views to build — pass at least one --view-html, or build into a bundle that already has a views.json.
  • A view tab 404s — its views/NN-*.html was removed; rebuild it with --view-html.
  • Empty or trivial input — if there is no real target to explain, ask the user for the actual document / repo / PR and produce the YAML first; do not emit an empty explainer.

Try it with the bundled sample

python scripts/build_html.py \
  --bundle ./sample-bundle \
  --core ../generate-explainer-yaml/references/sample-core.yaml \
  --view ../generate-explainer-yaml/references/sample-view.yaml \
  --quiz ../generate-explainer-yaml/references/sample-quiz.yaml \
  --prompts references/sample-prompts.json \
  --view-html "エンジニア=references/sample-iframe.html" \
  --view-html "クイズ=references/sample-quiz-iframe.html"

python scripts/validate_html.py ./sample-bundle/index.html ./sample-bundle/views/*.html

Open sample-bundle/index.html (Firefox / served). The right pane shows an engineer-oriented PR explanation (worktree + reading order + review checklist) plus an interactive quiz tab; the left pane has the copyable "add a view" templates and the YAML viewers.

Reference material

  • references/html-generation-rules.md — safety + UI rules for view documents (src-loaded iframe, light default, #theme hash)
  • references/quiz-view-rules.md — rules for the interactive quiz view (per-type rendering, grading contract, audience adaptation)
  • references/output-bundle-structure.md — the bundle's structure and shell
  • references/prompt-template-patterns.md — "add a view" prompt templates + path placeholders
  • references/sample-iframe.html — sample iframe view (engineer)
  • references/sample-quiz-iframe.html — sample interactive quiz view
  • references/sample-prompts.json — sample "add a view" templates
  • references/examples.md — three worked reader views
  • agents/openai.yaml — portable description of this skill for non-Claude agents
  • YAML schema (produced by the sibling skill): ../generate-explainer-yaml/references/ (core-yaml-schema.md, view-yaml-schema.md, quiz-yaml-schema.md)

Keep looking

Skills are one crate of 628,853. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.