generate-schema

Generate Schema

Generate a CLAUDE.md that tells Claude Code how to operate a wikillm knowledge base.

Arguments

Parse from $ARGUMENTS. Expected format:

--name <kb-name> --mode <personal|project-solo|project-team> --path <vault-path> --features <comma-separated> [--domain <context>]

Example:

/wikillm:generate-schema --name research --mode personal --path ~/.wikillm/research --features slides,reports,web-clipper --domain "machine learning and AI research"

Instructions

Generate a complete CLAUDE.md file and write it to <vault-path>/CLAUDE.md. The file must be a coherent, natural-reading document — not a template with conditional blocks.

Required Sections

1. Title and Philosophy — what this KB is and how it works
2. Directory Structure — the vault layout (varies by mode and features)
3. Automation — scheduling rules (varies by mode)
4. Ingestion — how raw/ files are processed (reference /wikillm:ingest)
5. Query — how to answer questions (reference /wikillm:query)
6. Wiki Conventions — article format, naming, linking, frontmatter
7. Index System — INDEX.md, TAGS.md, SOURCES.md, RECENT.md, LOG.md formats
8. Rules — immutability of raw/, wiki ownership, commit workflow

Mode Variations

Personal:

• Philosophy: "This is your brain extension — a persistent, compounding knowledge artifact."
• Tone: informal, personal notes
• Automation: "Scheduled tasks run via Claude Desktop" with user's chosen frequency
• Query focus: general knowledge synthesis
• Commit rules: solo, push freely

Project Solo:

• Philosophy: "This knowledge base is your project's compiled understanding."
• Tone: technical, architectural
• Automation: same as personal
• Query focus: "Why is the code like this?", architecture decisions, domain knowledge
• Commit rules: solo, push freely

Project Team:

• Philosophy: "This is the team's shared knowledge base — compiled understanding that everyone reads."
• Tone: clear enough for any team member
• Automation: "Manual only — run /wikillm:ingest after adding sources. No scheduled automation in team mode."
• Query focus: onboarding context, architecture decisions, "ask the wiki, not Steve"
• Commit rules: attribute commits, coordinate pushes, pull before ingesting

Feature Variations

Only include sections for enabled features:

• slides: add outputs/slides/ to directory structure, add a "Slide Generation" section referencing /wikillm:marp-cli
• reports: add outputs/reports/ to directory structure, add a "Reports" section
• visualizations: add outputs/visualizations/ to directory structure, add a "Visualizations" section describing matplotlib/chart workflow
• web-clipper: add a "Web Clipper Pipeline" section documenting Obsidian Web Clipper → raw/ flow

If no output features are enabled, omit the outputs/ directory entirely.

Domain Context

If --domain is provided, use it to:

• Frame the philosophy section around the domain
• Suggest relevant tags in the conventions section (e.g., for "machine learning": suggest tags like transformers, attention, fine-tuning, rag, etc.)
• Include domain-relevant example queries in the Query section
• Set the right vocabulary level for the wiki tone

Constants (always include)

These sections are the same regardless of mode:

Wiki conventions:

- Filenames: kebab-case (e.g., `attention-mechanism.md`)
- Links: always use [[wikilinks]] for cross-references
- Frontmatter: YAML on every page (created, updated, tags, sources)
- One concept per article
- Update existing articles over creating near-duplicates
- Mark uncertainty with [unverified]

Article format template:

---
created: YYYY-MM-DD
updated: YYYY-MM-DD
tags: [tag1, tag2]
sources: [raw/source-file.ext]
---

# Article Title

Brief summary (2-3 sentences).

## Content

Main content organized with headers.

## Related

- [[Related Article 1]]
- [[Related Article 2]]

## Sources

- `raw/source-file.ext` — what was drawn from this source

Index formats:

• INDEX.md: content catalog, every page with link + one-line summary, by category
• TAGS.md: all tags with links to articles using each tag
• SOURCES.md: maps each raw/ file to wiki articles it informed
• RECENT.md: last 20 changed articles
• LOG.md: chronological record with format ## [YYYY-MM-DD HH:MM] operation | Title

Rules:

• raw/ is immutable — never delete or edit source files
• Wiki is LLM-owned — the human reads, the LLM writes
• Always update indices and log after wiki changes
• Always git commit + push after every operation batch
• Use [[wikilinks]] for all cross-references

Keep CLAUDE.md Lean

The generated CLAUDE.md is loaded into every Claude Code session's context whenever the vault is in scope. Long CLAUDE.md files eat budget that should be spent on actual work. Aim for under 200 lines, and put the most critical information in the first 10-15 lines so a fast-path session can decide how to proceed without reading the rest.

Lead with a "Query-critical summary" block — 4-6 bullets covering facts a query session absolutely must know before touching the vault. These should be the things that, if ignored, would cause the session to produce a wrong answer or edit the wrong file:

• Where the wiki lives (wiki/) vs where raw sources live (raw/)
• That raw/ is immutable input, wiki/ is the compiled synthesis to read
• That all cross-references use [[wikilinks]] and YAML frontmatter is required
• That indices live in wiki/_index/ and must be updated after any wiki change
• Any domain-specific gotcha unique to this KB (e.g., "this KB covers third-party libraries only, not project code")
• The canonical entry point for queries (INDEX.md → Read, then obsidian backlinks to catch hubs)

Example opening:

# {{KB name}}

**Query-critical summary:**

- Compiled knowledge lives in `wiki/`. Source material lives in `raw/` (immutable — never edit).
- All articles use `[[wikilinks]]` for cross-references and YAML frontmatter (`created`, `updated`, `tags`, `sources`).
- Indices (INDEX, TAGS, SOURCES, RECENT, LOG) live in `wiki/_index/` and must be updated after any wiki write.
- For queries: start from `wiki/_index/INDEX.md`, Read the pinned articles directly, then run `obsidian backlinks` on each to catch hub articles. Never pull facts from `raw/`.
- {{domain-specific gotcha, if any}}

## Philosophy
...

The rest of the CLAUDE.md (directory structure, automation, conventions, rules) follows as reference material. The query-critical block lets sessions that just need to answer a question short-circuit reading the whole file.

Output

Write the generated CLAUDE.md to <vault-path>/CLAUDE.md. Confirm to the user: "Generated CLAUDE.md at ."