ai-course-glossary

Building an AI Course Glossary

This skill helps you produce a curated terminology glossary for an AI course hosted on Skilljar (e.g. the free courses at anthropic.skilljar.com — Claude 101, Claude Code, MCP Intro, AI Fluency series, etc.) or a similar third-party LMS. The output is a single JSON file shaped exactly like the per-language dictionaries the SkillBridge Chrome extension ships at src/data/<lang>.json, so contributions flow naturally into the open-source upstream and the same glossary doubles as a standalone reference card.

When to Use

Activate when the user asks any of:

• "Build a glossary for the Claude 101 course in Italian"
• "I want to add Polish as a premium language to SkillBridge"
• "Translate the technical terminology of this Skilljar course to "
• "Generate a vocabulary list for https://anthropic.skilljar.com/courses/..."
• "Help me localize this AI course"

Skip when:

• The user wants a generic translator (no glossary structure) — point them at a translation skill instead
• The content isn't AI/ML training — terminology curation has no value on cooking or marketing courses
• The platform has native i18n (Coursera AI tracks since 2026, Microsoft Learn, Google AI courses) — those handle localization themselves; building a parallel glossary duplicates work

Before You Start

Confirm three things with the user:

1. Source URL: a specific Skilljar course page (or list of pages). Without a URL we can't extract real terminology — refuse to hallucinate a glossary from generic AI knowledge alone, because the value of this output is fidelity to the course's actual phrasing, not a generic dictionary.
2. Target language: an ISO 639-1 code (it, fr, es, de, pt-BR, ja, ko, zh-CN, zh-TW, ru, vi, pl, ar, hi, nl, tr, id, th, he, bn, etc.). If the user names a country, ask for the language explicitly — Spanish-vs-Catalan, Mandarin-vs-Cantonese, Brazilian-vs-European Portuguese all matter.
3. Output destination: standalone JSON file, or a SkillBridge PR. The schema is the same; only the file path and the surrounding commit message differ.

Output Schema

Match the SkillBridge v3.5.34 dictionary schema exactly. The schema is intentionally section-rigid — every premium-tier dictionary in the SkillBridge repo ships the same 22 sections with identical English keys, and scripts/check-dict-coverage.js enforces this in CI. A glossary that misses or renames sections will fail validation and won't merge. See references/skillbridge-schema.md for the full section list with key counts.

Top-level shape:

{
  "_meta": {
    "lang": "<iso-code>",
    "langName": "<native-name>",
    "version": "<skillbridge-version>",
    "lastUpdated": "<YYYY-MM-DD>",
    "translation_provenance": "<honest-source-disclosure>"
  },
  "ui": { "<English>": "<translated>", ... },
  "catalog": { ... },
  "faq": { ... },
  "claude101": { ... },
  "claudeCode": { ... },
  "agentSkills": { ... },
  "claudeCowork": { ... },
  "subagents": { ... },
  "aiFluency": { ... },
  "common": { ... },
  "mcpIntro": { ... },
  "mcpAdvanced": { ... },
  "aiCapabilities": { ... },
  "claudeAPI": { ... },
  "aiFluencyEdu": { ... },
  "aiFluencyStudent": { ... },
  "aiFluencyNonprofit": { ... },
  "cloudDeployment": { ... },
  "extendedThinking": { ... },
  "teachingAI": { ... },
  "_protected": { ... },
  "exam_ui": { ... }
}

Two of these sections are non-translation:

• _meta — file metadata. Always present. translation_provenance should honestly state whether the translation was produced by a native speaker, derived from a sibling Romance/CJK language, or AI-translated and pending review. SkillBridge surfaces this in code review.
• _protected — brand names (Anthropic, Claude, Cowork, Dispatch, Computer Use, Subagent, etc.) that must NEVER be translated and that Google Translate must be auto-corrected against if it tries. Copy verbatim from any existing dictionary in the upstream repo — these are stable across all languages.

The remaining 20 sections are full English-key → target-language translations.

Workflow

1. Bootstrap from an existing dictionary. Don't start from a blank file. WebFetch one of src/data/es.json, src/data/fr.json, or src/data/ja.json from github.com/heznpc/skillBridge (pick the closest Romance / CJK relative) and use it as a scaffold. This guarantees structural compatibility with check-dict-coverage.js and avoids the easy mistake of missing a section.

2. Verify the source URL is live AI content. Open the user-provided course URL with WebFetch. Confirm:

   • The page is on a recognized Skilljar tenant (the URL host ends in .skilljar.com, or it's a Coursera AI track, etc.)
   • The page actually surfaces AI terminology (prompt, model, agent, MCP, Claude, GPT, RAG, fine-tuning, embedding, etc.) — not a generic corporate training page that happens to be on Skilljar
   • The course slug matches a known section name in the SkillBridge schema (see references/skillbridge-schema.md). If the slug is new, propose a new section name following the existing convention (camelCase derived from the course title).

3. Extract terminology. Walk the course page text and pull out:

   • UI strings: navigation labels, button text, lesson titles, breadcrumbs. These live in the ui section and are mostly shared across courses.
   • Course-specific terminology: technical terms the course teaches (e.g. for Claude 101: "Constitutional AI", "Harmlessness", "Helpfulness", "RLHF"; for MCP Intro: "Server", "Client", "Tool", "Resource", "Sampling"). These live in the section named after the course slug.
   • Headings, summaries, takeaways: navigational text that helps the learner orient. These live in common.

4. Translate. For each extracted English term, produce the target-language translation following three rules:

   • Brand names stay English — copy from _protected. "Claude" never becomes "클로드"; "Anthropic" never becomes "Anthropique". The whole point of _protected is to undo Google Translate's tendency to phonetic-transliterate brand names.
   • Technical terms use the field-standard convention in the target language. For Korean, "prompt" is 프롬프트 (loanword), not 유도질문. For Italian, "fine-tuning" is fine-tuning (loanword), not messa a punto fine. When in doubt, check what HuggingFace / OpenAI / DeepMind localized docs use.
   • Natural-language strings are translated normally — full sentences, not word-for-word.

5. Disclose provenance honestly. In _meta.translation_provenance, write one of:

   • "v1 — native speaker, hand-curated <YYYY-MM-DD>" if a native speaker reviewed every entry
   • "v1 — derived from src/data/<source>.json via <source-lang>→<target-lang> regex transformation (proximity ~N%). Native review welcome." if you bootstrapped from a sibling language (the SkillBridge Italian dictionary v3.5.34 used this exact pattern from Spanish)
   • "v1 — AI-translated, pending native review" if Claude generated the translations cold without a sibling language to lean on

6. Validate locally before handing the file back. Two checks:

   • Open the existing SkillBridge dictionaries via WebFetch and confirm the section list matches exactly (count + names).
   • Confirm _protected is a verbatim copy of the upstream _protected section (don't translate brand names by accident).

7. Hand off. Two delivery modes:

   • Standalone glossary: just hand the user the JSON file at <lang>.json. They use it as a reference card.
   • SkillBridge PR-ready: file at src/data/<lang>.json plus a commit message like feat(i18n): promote <lang> to Premium tier with curated dictionary v1. The user clones the SkillBridge repo, drops the file in, opens a PR. The PR will need to:
     • Add the language to PREMIUM_LANGUAGES in src/lib/constants.js
     • Remove it from the Standard list (the spread of PREMIUM_LANGUAGES already includes it)
     • Update _locales/<lang>/messages.json extDescription
     • Update README.md and store-assets/STORE_LISTING*.md Premium tables
     • Bump the PREMIUM_LANGUAGES count assertion in tests/constants.test.js and tests/translator.test.js See examples/skillbridge-pr-checklist.md for the full upstream-PR mechanics.

Common Pitfalls

• Hallucinating a course that doesn't exist. Always fetch the user-provided URL first. If the URL is unreachable, ask the user to verify it before producing anything. A glossary for a non-existent course is worse than no glossary.

• Translating brand names. Triple-check _protected. "Claude" is not "Claudio" in Italian, not "クロード" in Japanese.

• Missing a section. SkillBridge's check-dict-coverage.js will reject a dictionary with even one missing section. Always use an existing dictionary as the structural scaffold.

• Inventing terminology that doesn't match what the field uses. If you're translating to a language you're not deeply familiar with, prefer loanwords for technical AI terms (the field-standard in most non-English locales is loanword adoption, e.g. Korean "프롬프트 엔지니어링", Italian "prompt engineering" verbatim).

• Adding new sections. Don't. SkillBridge's section schema is locked at v3.5.34. If a new Anthropic Academy course appears (the academy-courses-drift watcher in the SkillBridge repo will catch this within 12 hours of a course launch), add it as a new section in all premium dictionaries at once via a coordinated PR, not just yours.

• Burning Puter.js quota during extraction. This skill runs in Claude Code; it uses Claude Code's compute. The SkillBridge extension uses Puter.js for runtime translation. Don't try to call Puter.js from this skill — wrong surface.

References

• references/skillbridge-schema.md — full section list with current key counts, regenerated whenever the upstream schema changes
• references/skillbridge-pr-checklist.md — step-by-step for opening a PR against heznpc/skillBridge with a new language dictionary
• examples/ — input URL + output JSON pairs for the three most common cases (Italian/French/Spanish derivations from existing dictionaries)
• Upstream SkillBridge repo: https://github.com/heznpc/skillBridge
• Upstream POSITIONING (read before submitting a major change): https://github.com/heznpc/skillBridge/blob/main/POSITIONING.md

Related Skills

This skill is the content counterpart to SkillBridge's runtime surface. The SkillBridge Chrome extension translates Skilljar pages live and ships with a curated dictionary per premium language; this skill helps you produce that dictionary for a new language. The two are independent — you can use this glossary skill without ever installing the extension, and the extension works without this skill (its 11 premium dictionaries are bundled in the extension package).