Bulk Generate Concepts

Process

1. Clarify scope - Subject, grades, curriculum standard (e.g., German Rahmenlehrplan)

2. Create feature branch - Before any work:

   git fetch origin dev
   git checkout -b concepts/{subject}-grade{N} origin/dev
   

   Branch naming: concepts/{subject}-grade{N} (e.g., concepts/german-grade4)

3. Research curriculum ONCE - Use official sources (ministry sites). Extract:

   • Learning objectives per concept
   • Age-appropriate expectations
   • Difficulty progression
   • Key terminology

   Save to docs/research/{subject}-curriculum-research.md with:

   • Concept list (IDs, titles, grades, scope)
   • Extracted curriculum details per concept
   • Source URLs

   Get user approval on concept list.

4. Spawn parallel agents - One Task tool call per concept:

   Create concept: {subject}/{concept-id} (Grade {grade})
   
   1. Read `docs/concept-rules.md` first
   2. Use write-concept skill
   3. Reference `docs/{subject}-curriculum-research.md` for curriculum context
   4. **SKIP RESEARCH**: All content is provided!
   
   Scope: {one-sentence scope description}
   

   Do NOT paste curriculum content into agent prompts. Agents read the research doc themselves. Keep prompts minimal.

5. Verify files exist - Check that concept files were actually created. Agents may report success without writing files.

6. Validate - Use bun run check:concepts to efficiently check the generated concepts. If failures, resume agents with fixes.

7. Commit and create PR - Stage, commit, push, and open PR against dev:

   git add content/subjects/{subject}/ packages/frontend/public/locales/
   git commit -m "feat(content): add {Subject} Grade {N} concepts (curriculum aligned)"
   git push -u origin concepts/{subject}-grade{N}
   gh pr create --base dev --title "feat(content): {Subject} Grade {N} concepts" --body "Add {N} new concepts for Grade {N} {Subject} based on {curriculum}."
   

8. Cleanup - Delete docs/{subject}-curriculum-research.md after PR is created.

Constraints

• Max 12-14 concepts per grade (consolidate if curriculum suggests more)
• Subagents read write-concept skill themselves - don't duplicate rules
• Research happens ONCE in parent, not per-agent (saves tokens)

Common Rationalizations